From 15b86e85334eaa98910a15ea84580dff7da1e9c5 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 3 Jun 2024 10:14:57 +0800
Subject: [PATCH 01/51] experience pool

---
 config/config2.example.yaml        |  3 +++
 examples/exp_pool/manager.py       | 21 ++++++++++++++++++++
 metagpt/config2.py                 |  4 ++++
 metagpt/configs/exp_pool_config.py |  6 ++++++
 metagpt/exp_pool/__init__.py       |  0
 metagpt/exp_pool/decorator.py      |  4 ++++
 metagpt/exp_pool/manager.py        | 32 ++++++++++++++++++++++++++++++
 metagpt/exp_pool/schema.py         | 25 +++++++++++++++++++++++
 metagpt/utils/reflection.py        |  9 +++++++++
 metagpt/utils/token_counter.py     |  2 ++
 10 files changed, 106 insertions(+)
 create mode 100644 examples/exp_pool/manager.py
 create mode 100644 metagpt/configs/exp_pool_config.py
 create mode 100644 metagpt/exp_pool/__init__.py
 create mode 100644 metagpt/exp_pool/decorator.py
 create mode 100644 metagpt/exp_pool/manager.py
 create mode 100644 metagpt/exp_pool/schema.py

diff --git a/config/config2.example.yaml b/config/config2.example.yaml
index f1158775b..c5ca6e767 100644
--- a/config/config2.example.yaml
+++ b/config/config2.example.yaml
@@ -74,6 +74,9 @@ s3:
   secure: false
   bucket: "test"
 
+experience_pool:
+  enable_read: false
+  enable_write: false
 
 azure_tts_subscription_key: "YOUR_SUBSCRIPTION_KEY"
 azure_tts_region: "eastus"
diff --git a/examples/exp_pool/manager.py b/examples/exp_pool/manager.py
new file mode 100644
index 000000000..f5766f9a5
--- /dev/null
+++ b/examples/exp_pool/manager.py
@@ -0,0 +1,21 @@
+from metagpt.exp_pool.manager import ExperiencePoolManager
+from metagpt.exp_pool.schema import Experience
+from pprint import pprint
+import asyncio
+# import logging
+# logging.basicConfig(level=logging.DEBUG)
+
+async def main():
+    req = "2048 game"
+    exp = Experience(req=req, resp="python code")
+
+    manager = ExperiencePoolManager()
+
+    # pprint(manager.storage.get())
+    # manager.create_exp(exp)
+    result = await manager.query_exp(req)
+    print(result)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
\ No newline at end of file
diff --git a/metagpt/config2.py b/metagpt/config2.py
index 8c61fdbf2..3f8930401 100644
--- a/metagpt/config2.py
+++ b/metagpt/config2.py
@@ -21,6 +21,7 @@ from metagpt.configs.search_config import SearchConfig
 from metagpt.configs.workspace_config import WorkspaceConfig
 from metagpt.const import CONFIG_ROOT, METAGPT_ROOT
 from metagpt.utils.yaml_model import YamlModel
+from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 
 
 class CLIParams(BaseModel):
@@ -67,6 +68,9 @@ class Config(CLIParams, YamlModel):
     enable_longterm_memory: bool = False
     code_review_k_times: int = 2
 
+    # Experience Pool Parameters
+    experience_pool: Optional[ExperiencePoolConfig] = None
+
     # Will be removed in the future
     metagpt_tti_url: str = ""
     language: str = "English"
diff --git a/metagpt/configs/exp_pool_config.py b/metagpt/configs/exp_pool_config.py
new file mode 100644
index 000000000..f7312d2de
--- /dev/null
+++ b/metagpt/configs/exp_pool_config.py
@@ -0,0 +1,6 @@
+from metagpt.utils.yaml_model import YamlModel
+
+
+class ExperiencePoolConfig(YamlModel):
+    enable_read: bool = False
+    enable_write: bool = False
diff --git a/metagpt/exp_pool/__init__.py b/metagpt/exp_pool/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
new file mode 100644
index 000000000..6629e8377
--- /dev/null
+++ b/metagpt/exp_pool/decorator.py
@@ -0,0 +1,4 @@
+
+
+def exp_cache(func):
+    pass
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
new file mode 100644
index 000000000..c32073a9f
--- /dev/null
+++ b/metagpt/exp_pool/manager.py
@@ -0,0 +1,32 @@
+from pydantic import BaseModel, ConfigDict
+from metagpt.exp_pool.schema import Experience
+import uuid
+import chromadb
+from chromadb import Collection, QueryResult
+from typing import Optional
+from metagpt.rag.engines import SimpleEngine
+from metagpt.rag.schema import ChromaRetrieverConfig
+
+
+class ExperiencePoolManager(BaseModel):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self._storage = None
+
+    @property
+    def storage(self) -> SimpleEngine:
+        if self._storage is None:
+            self._storage = SimpleEngine.from_objs(retriever_configs=[ChromaRetrieverConfig(collection_name="experience_pool", persist_path="./chroma_data")])
+        return self._storage
+    
+    def create_exp(self, exp: Experience):
+        self.storage.add_objs([exp])
+    
+    async def query_exp(self, req: str) -> list[Experience]:
+        nodes = await self.storage.aretrieve(req)
+        exps = [node.metadata["obj"] for node in nodes]
+
+        return exps
+
+    
+
diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
new file mode 100644
index 000000000..359268612
--- /dev/null
+++ b/metagpt/exp_pool/schema.py
@@ -0,0 +1,25 @@
+from pydantic import BaseModel, Field
+from llama_index.core.schema import TextNode
+
+
+class Experience(BaseModel):
+    req: str = Field(..., description="")
+    resp: str = Field(..., description="")
+
+    def rag_key(self):
+        return self.req
+
+
+class ExperienceNodeMetadata(BaseModel):
+    """Metadata of ExperienceNode."""
+
+    resp: str = Field(..., description="")
+
+
+class ExperienceNode(TextNode):
+    """ExperienceNode for RAG."""
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.excluded_llm_metadata_keys = list(ExperienceNodeMetadata.model_fields.keys())
+        self.excluded_embed_metadata_keys = self.excluded_llm_metadata_keys
diff --git a/metagpt/utils/reflection.py b/metagpt/utils/reflection.py
index 8b8237ae7..688831f06 100644
--- a/metagpt/utils/reflection.py
+++ b/metagpt/utils/reflection.py
@@ -1,4 +1,5 @@
 """class tools, including method inspection, class attributes, inheritance relationships, etc."""
+import inspect
 
 
 def check_methods(C, *methods):
@@ -16,3 +17,11 @@ def check_methods(C, *methods):
         else:
             return NotImplemented
     return True
+
+
+def get_func_full_name(func, *args) -> str:
+    if inspect.ismethod(func) or (inspect.isfunction(func) and "self" in inspect.signature(func).parameters):
+        cls_name = args[0].__class__.__name__
+        return f"{func.__module__}.{cls_name}.{func.__name__}"
+    
+    return f"{func.__module__}.{func.__name__}"
diff --git a/metagpt/utils/token_counter.py b/metagpt/utils/token_counter.py
index 0ba2daa89..496842a2d 100644
--- a/metagpt/utils/token_counter.py
+++ b/metagpt/utils/token_counter.py
@@ -150,6 +150,8 @@ TOKEN_MAX = {
     "gpt-4-1106-preview": 128000,
     "gpt-4-vision-preview": 128000,
     "gpt-4-1106-vision-preview": 128000,
+    "gpt-4-turbo": 128000,
+    "gpt-4o": 128000,
     "gpt-4": 8192,
     "gpt-4-0613": 8192,
     "gpt-4-32k": 32768,

From 471310f3b3e879ea269d2941f525c83e04b55938 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 10:28:39 +0800
Subject: [PATCH 02/51] experiment pool init

---
 .gitignore                               |   1 +
 config/config2.example.yaml              |   6 +-
 examples/exp_pool/decorator.py           |  26 ++++++
 examples/exp_pool/manager.py             |  21 -----
 examples/exp_pool/simple.py              |  29 ++++++
 metagpt/config2.py                       |   4 +-
 metagpt/configs/exp_pool_config.py       |   6 +-
 metagpt/exp_pool/__init__.py             |   6 ++
 metagpt/exp_pool/decorator.py            |  56 ++++++++++-
 metagpt/exp_pool/manager.py              | 113 +++++++++++++++++++----
 metagpt/exp_pool/schema.py               |  40 +++++++-
 metagpt/rag/retrievers/bm25_retriever.py |   2 +-
 metagpt/utils/file.py                    |   1 -
 metagpt/utils/reflection.py              |   2 +-
 14 files changed, 258 insertions(+), 55 deletions(-)
 create mode 100644 examples/exp_pool/decorator.py
 delete mode 100644 examples/exp_pool/manager.py
 create mode 100644 examples/exp_pool/simple.py

diff --git a/.gitignore b/.gitignore
index aa5edd74a..7c64829ad 100644
--- a/.gitignore
+++ b/.gitignore
@@ -162,6 +162,7 @@ examples/graph_store.json
 examples/image__vector_store.json
 examples/index_store.json
 .chroma
+.chroma_exp_data
 *~$*
 workspace/*
 tmp
diff --git a/config/config2.example.yaml b/config/config2.example.yaml
index c5ca6e767..c7b2cae2c 100644
--- a/config/config2.example.yaml
+++ b/config/config2.example.yaml
@@ -74,9 +74,9 @@ s3:
   secure: false
   bucket: "test"
 
-experience_pool:
-  enable_read: false
-  enable_write: false
+exp_pool:
+  enable_read: true
+  enable_write: true
 
 azure_tts_subscription_key: "YOUR_SUBSCRIPTION_KEY"
 azure_tts_region: "eastus"
diff --git a/examples/exp_pool/decorator.py b/examples/exp_pool/decorator.py
new file mode 100644
index 000000000..2f6397f80
--- /dev/null
+++ b/examples/exp_pool/decorator.py
@@ -0,0 +1,26 @@
+"""Decorator example of experience pool."""
+
+import asyncio
+import uuid
+
+from metagpt.exp_pool import exp_cache, exp_manager
+from metagpt.logs import logger
+
+
+@exp_cache
+async def produce(req):
+    return f"{req} {uuid.uuid4().hex}"
+
+
+async def main():
+    req = "Water"
+
+    resp = await produce(req)
+    logger.info(f"The resp of `produce{req}` is: {resp}")
+
+    exps = await exp_manager.query_exps(req)
+    logger.info(f"Find experiences: {exps}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/examples/exp_pool/manager.py b/examples/exp_pool/manager.py
deleted file mode 100644
index f5766f9a5..000000000
--- a/examples/exp_pool/manager.py
+++ /dev/null
@@ -1,21 +0,0 @@
-from metagpt.exp_pool.manager import ExperiencePoolManager
-from metagpt.exp_pool.schema import Experience
-from pprint import pprint
-import asyncio
-# import logging
-# logging.basicConfig(level=logging.DEBUG)
-
-async def main():
-    req = "2048 game"
-    exp = Experience(req=req, resp="python code")
-
-    manager = ExperiencePoolManager()
-
-    # pprint(manager.storage.get())
-    # manager.create_exp(exp)
-    result = await manager.query_exp(req)
-    print(result)
-
-
-if __name__ == "__main__":
-    asyncio.run(main())
\ No newline at end of file
diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/simple.py
new file mode 100644
index 000000000..bc20fbcdd
--- /dev/null
+++ b/examples/exp_pool/simple.py
@@ -0,0 +1,29 @@
+"""Simple example of experience pool."""
+
+import asyncio
+
+from metagpt.exp_pool import exp_manager
+from metagpt.exp_pool.schema import EntryType, Experience
+from metagpt.logs import logger
+
+
+async def main():
+    req = "Simple task."
+
+    # 1. Find experiences.
+    exps = await exp_manager.query_exps(req)
+    if exps:
+        logger.info(f"Experiences already exist for the request `{req}`: {exps}")
+        return
+
+    # 2. Create a new experience if none exist
+    exp_manager.create_exp(Experience(req=req, resp="Simple echo.", entry_type=EntryType.MANUAL))
+    logger.info(f"New experience created for the request `{req}`.")
+
+    # 3. Find again
+    exps = await exp_manager.query_exps(req)
+    logger.info(f"Updated experiences: {exps}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/metagpt/config2.py b/metagpt/config2.py
index 6f5a1add6..6588a6036 100644
--- a/metagpt/config2.py
+++ b/metagpt/config2.py
@@ -13,6 +13,7 @@ from pydantic import BaseModel, model_validator
 
 from metagpt.configs.browser_config import BrowserConfig
 from metagpt.configs.embedding_config import EmbeddingConfig
+from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 from metagpt.configs.llm_config import LLMConfig, LLMType
 from metagpt.configs.mermaid_config import MermaidConfig
 from metagpt.configs.redis_config import RedisConfig
@@ -22,7 +23,6 @@ from metagpt.configs.search_config import SearchConfig
 from metagpt.configs.workspace_config import WorkspaceConfig
 from metagpt.const import CONFIG_ROOT, METAGPT_ROOT
 from metagpt.utils.yaml_model import YamlModel
-from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 
 
 class CLIParams(BaseModel):
@@ -73,7 +73,7 @@ class Config(CLIParams, YamlModel):
     code_review_k_times: int = 2
 
     # Experience Pool Parameters
-    experience_pool: Optional[ExperiencePoolConfig] = None
+    exp_pool: ExperiencePoolConfig = ExperiencePoolConfig()
 
     # Will be removed in the future
     metagpt_tti_url: str = ""
diff --git a/metagpt/configs/exp_pool_config.py b/metagpt/configs/exp_pool_config.py
index f7312d2de..3f86173c1 100644
--- a/metagpt/configs/exp_pool_config.py
+++ b/metagpt/configs/exp_pool_config.py
@@ -1,6 +1,8 @@
+from pydantic import Field
+
 from metagpt.utils.yaml_model import YamlModel
 
 
 class ExperiencePoolConfig(YamlModel):
-    enable_read: bool = False
-    enable_write: bool = False
+    enable_read: bool = Field(default=True, description="Enable to read from experience pool.")
+    enable_write: bool = Field(default=True, description="Enable to write to experience pool.")
diff --git a/metagpt/exp_pool/__init__.py b/metagpt/exp_pool/__init__.py
index e69de29bb..aeeb94b38 100644
--- a/metagpt/exp_pool/__init__.py
+++ b/metagpt/exp_pool/__init__.py
@@ -0,0 +1,6 @@
+"""Experience pool init."""
+
+from metagpt.exp_pool.manager import exp_manager
+from metagpt.exp_pool.decorator import exp_cache
+
+__all__ = ["exp_manager", "exp_cache"]
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 6629e8377..1d691b8f3 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -1,4 +1,56 @@
+"""Experience Decorator."""
+
+import asyncio
+import functools
+from typing import Any, Callable, Optional, TypeVar
+
+from metagpt.exp_pool.manager import exp_manager
+from metagpt.exp_pool.schema import Experience
+from metagpt.utils.async_helper import NestAsyncio
+
+ReturnType = TypeVar("ReturnType")
 
 
-def exp_cache(func):
-    pass
+def exp_cache(_func: Optional[Callable[..., ReturnType]] = None):
+    """Decorator to check for a perfect experience and returns it if exists.
+
+    Otherwise, it executes the function, save the result as a new experience, and returns the result.
+
+    This can be applied to both synchronous and asynchronous functions.
+    """
+
+    def decorator(func: Callable[..., ReturnType]) -> Callable[..., ReturnType]:
+        @functools.wraps(func)
+        async def get_or_create(args: Any, kwargs: Any, is_async: bool) -> ReturnType:
+            """Attempts to retrieve a cached experience or creates one if not found."""
+
+            req = f"{func.__name__}_{args}_{kwargs}"
+            exps = await exp_manager.query_exps(req)
+            if perfect_exp := exp_manager.extract_one_perfect_exp(exps):
+                return perfect_exp
+
+            if is_async:
+                result = await func(*args, **kwargs)
+            else:
+                result = func(*args, **kwargs)
+
+            exp_manager.create_exp(Experience(req=req, resp=result))
+
+            return result
+
+        def sync_wrapper(*args: Any, **kwargs: Any) -> ReturnType:
+            NestAsyncio.apply_once()
+            return asyncio.get_event_loop().run_until_complete(get_or_create(args, kwargs, is_async=False))
+
+        async def async_wrapper(*args: Any, **kwargs: Any) -> ReturnType:
+            return await get_or_create(args, kwargs, is_async=True)
+
+        if asyncio.iscoroutinefunction(func):
+            return async_wrapper
+        else:
+            return sync_wrapper
+
+    if _func is None:
+        return decorator
+    else:
+        return decorator(_func)
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index c32073a9f..4bc566104 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -1,32 +1,105 @@
-from pydantic import BaseModel, ConfigDict
-from metagpt.exp_pool.schema import Experience
-import uuid
-import chromadb
-from chromadb import Collection, QueryResult
+"""Experience Manager."""
+
 from typing import Optional
+
+from pydantic import BaseModel, ConfigDict, model_validator
+
+from metagpt.config2 import Config, config
+from metagpt.exp_pool.schema import MAX_SCORE, Experience
 from metagpt.rag.engines import SimpleEngine
-from metagpt.rag.schema import ChromaRetrieverConfig
+from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
 
 
-class ExperiencePoolManager(BaseModel):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self._storage = None
+class ExperienceManager(BaseModel):
+    """ExperienceManager manages the lifecycle of experiences, including CRUD and optimization.
+
+    Attributes:
+        config (Config): Configuration for managing experiences.
+        storage (SimpleEngine): Engine to handle the storage and retrieval of experiences.
+    """
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    config: Config = config
+    storage: SimpleEngine = None
+
+    @model_validator(mode="after")
+    def initialize(self):
+        if self.storage is None:
+            self.storage = SimpleEngine.from_objs(
+                retriever_configs=[
+                    ChromaRetrieverConfig(collection_name="experience_pool", persist_path=".chroma_exp_data")
+                ],
+                ranker_configs=[LLMRankerConfig()],
+            )
+        return self
 
-    @property
-    def storage(self) -> SimpleEngine:
-        if self._storage is None:
-            self._storage = SimpleEngine.from_objs(retriever_configs=[ChromaRetrieverConfig(collection_name="experience_pool", persist_path="./chroma_data")])
-        return self._storage
-    
     def create_exp(self, exp: Experience):
+        """Adds an experience to the storage if writing is enabled.
+
+        Args:
+            exp (Experience): The experience to add.
+        """
+        if not self.config.exp_pool.enable_write:
+            return
+
         self.storage.add_objs([exp])
-    
-    async def query_exp(self, req: str) -> list[Experience]:
+
+    async def query_exps(self, req: str, tag: str = "") -> list[Experience]:
+        """Retrieves and filters experiences.
+
+        Args:
+            req (str): The query string to retrieve experiences.
+            tag (str): Optional tag to filter the experiences by.
+
+        Returns:
+            list[Experience]: A list of experiences that match the args.
+        """
+        if not self.config.exp_pool.enable_read:
+            return []
+
         nodes = await self.storage.aretrieve(req)
-        exps = [node.metadata["obj"] for node in nodes]
+        exps: list[Experience] = [node.metadata["obj"] for node in nodes]
+
+        # TODO: filter by metadata
+        if tag:
+            exps = [exp for exp in exps if exp.tag == tag]
 
         return exps
 
-    
+    def extract_one_perfect_exp(self, exps: list[Experience]) -> Optional[Experience]:
+        """Extracts the first 'perfect' experience from a list of experiences.
 
+        Args:
+            exps (list[Experience]): The experiences to evaluate.
+
+        Returns:
+            Optional[Experience]: The first perfect experience if found, otherwise None.
+        """
+        for exp in exps:
+            if self.is_perfect_exp(exp):
+                return exp
+
+        return None
+
+    @staticmethod
+    def is_perfect_exp(exp: Experience) -> bool:
+        """Determines if an experience is considered 'perfect'.
+
+        Args:
+            exp (Experience): The experience to evaluate.
+
+        Returns:
+            bool: True if the experience is manually entered, otherwise False.
+        """
+        if not exp:
+            return False
+
+        # TODO: need more metrics
+        if exp.metric and exp.metric.score == MAX_SCORE:
+            return True
+
+        return False
+
+
+exp_manager = ExperienceManager()
diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index 359268612..b51bc3c17 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -1,10 +1,46 @@
-from pydantic import BaseModel, Field
+"""Experience schema."""
+
+from enum import Enum
+from typing import Optional
+
 from llama_index.core.schema import TextNode
+from pydantic import BaseModel, Field
+
+MAX_SCORE = 10
+
+
+class ExperienceType(str, Enum):
+    """Experience Type."""
+
+    SUCCESS = "success"
+    FAILURE = "failure"
+    INSIGHT = "insight"
+
+
+class EntryType(Enum):
+    """Experience Entry Type."""
+
+    AUTOMATIC = "Automatic"
+    MANUAL = "Manual"
+
+
+class Metric(BaseModel):
+    """Experience Metric."""
+
+    time_cost: float = Field(default=0.000, description="Time cost, the unit is milliseconds.")
+    money_cost: float = Field(default=0.000, description="Money cost, the unit is US dollars.")
+    score: int = Field(default=1, description="Score, a value between 1 and 10.")
 
 
 class Experience(BaseModel):
+    """Experience."""
+
     req: str = Field(..., description="")
-    resp: str = Field(..., description="")
+    resp: str = Field(..., description="The type is string/json/code.")
+    metric: Optional[Metric] = Field(default=None, description="Metric.")
+    exp_type: ExperienceType = Field(default=ExperienceType.SUCCESS, description="The type of experience.")
+    entry_type: EntryType = Field(default=EntryType.AUTOMATIC, description="Type of entry: Manual or Automatic.")
+    tag: str = Field(default="", description="Tagging experience.")
 
     def rag_key(self):
         return self.req
diff --git a/metagpt/rag/retrievers/bm25_retriever.py b/metagpt/rag/retrievers/bm25_retriever.py
index 3b085cb73..dc75d87b0 100644
--- a/metagpt/rag/retrievers/bm25_retriever.py
+++ b/metagpt/rag/retrievers/bm25_retriever.py
@@ -46,4 +46,4 @@ class DynamicBM25Retriever(BM25Retriever):
     def persist(self, persist_dir: str, **kwargs) -> None:
         """Support persist."""
         if self._index:
-            self._index.storage_context.persist(persist_dir)
\ No newline at end of file
+            self._index.storage_context.persist(persist_dir)
diff --git a/metagpt/utils/file.py b/metagpt/utils/file.py
index a8ed482d9..8861f65dc 100644
--- a/metagpt/utils/file.py
+++ b/metagpt/utils/file.py
@@ -72,7 +72,6 @@ class File:
 
 
 class MemoryFileSystem(_MemoryFileSystem):
-
     @classmethod
     def _strip_protocol(cls, path):
         return super()._strip_protocol(str(path))
diff --git a/metagpt/utils/reflection.py b/metagpt/utils/reflection.py
index 688831f06..2683e5657 100644
--- a/metagpt/utils/reflection.py
+++ b/metagpt/utils/reflection.py
@@ -23,5 +23,5 @@ def get_func_full_name(func, *args) -> str:
     if inspect.ismethod(func) or (inspect.isfunction(func) and "self" in inspect.signature(func).parameters):
         cls_name = args[0].__class__.__name__
         return f"{func.__module__}.{cls_name}.{func.__name__}"
-    
+
     return f"{func.__module__}.{func.__name__}"

From 6d983908314622125033459cce16c401d47bc87d Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 11:33:34 +0800
Subject: [PATCH 03/51] experiment pool init

---
 examples/exp_pool/simple.py | 13 +++----------
 1 file changed, 3 insertions(+), 10 deletions(-)

diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/simple.py
index bc20fbcdd..608578519 100644
--- a/examples/exp_pool/simple.py
+++ b/examples/exp_pool/simple.py
@@ -9,20 +9,13 @@ from metagpt.logs import logger
 
 async def main():
     req = "Simple task."
+    resp = "Simple echo."
 
-    # 1. Find experiences.
-    exps = await exp_manager.query_exps(req)
-    if exps:
-        logger.info(f"Experiences already exist for the request `{req}`: {exps}")
-        return
-
-    # 2. Create a new experience if none exist
-    exp_manager.create_exp(Experience(req=req, resp="Simple echo.", entry_type=EntryType.MANUAL))
+    exp_manager.create_exp(Experience(req=req, resp=resp, entry_type=EntryType.MANUAL))
     logger.info(f"New experience created for the request `{req}`.")
 
-    # 3. Find again
     exps = await exp_manager.query_exps(req)
-    logger.info(f"Updated experiences: {exps}")
+    logger.info(f"Got experiences: {exps}")
 
 
 if __name__ == "__main__":

From 9f817bd59c254ae0b86b1b1ef9d5d57b6f63a44a Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 11:46:57 +0800
Subject: [PATCH 04/51] experiment pool init

---
 examples/exp_pool/simple.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/simple.py
index 608578519..f270824bf 100644
--- a/examples/exp_pool/simple.py
+++ b/examples/exp_pool/simple.py
@@ -10,8 +10,9 @@ from metagpt.logs import logger
 async def main():
     req = "Simple task."
     resp = "Simple echo."
+    exp = Experience(req=req, resp=resp, entry_type=EntryType.MANUAL)
 
-    exp_manager.create_exp(Experience(req=req, resp=resp, entry_type=EntryType.MANUAL))
+    exp_manager.create_exp(exp)
     logger.info(f"New experience created for the request `{req}`.")
 
     exps = await exp_manager.query_exps(req)

From d10881c0e441e0f37b1645c4ca890c30f2868da3 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 15:16:58 +0800
Subject: [PATCH 05/51] add trajectory schema

---
 metagpt/exp_pool/schema.py | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index b51bc3c17..e6ae4ee1d 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -32,6 +32,14 @@ class Metric(BaseModel):
     score: int = Field(default=1, description="Score, a value between 1 and 10.")
 
 
+class Trajectory(BaseModel):
+    """Experience Trajectory."""
+
+    plan: str = Field(default="", description="The plan.")
+    action: str = Field(default="", description="Action for the plan.")
+    observation: str = Field(default="", description="Output of the action.")
+
+
 class Experience(BaseModel):
     """Experience."""
 
@@ -41,6 +49,7 @@ class Experience(BaseModel):
     exp_type: ExperienceType = Field(default=ExperienceType.SUCCESS, description="The type of experience.")
     entry_type: EntryType = Field(default=EntryType.AUTOMATIC, description="Type of entry: Manual or Automatic.")
     tag: str = Field(default="", description="Tagging experience.")
+    traj: Optional[Trajectory] = Field(default=None, description="Trajectory.")
 
     def rag_key(self):
         return self.req

From 2eedc23a827acc892c4928bf14c7e1b99f081c59 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 22:08:40 +0800
Subject: [PATCH 06/51] add exp_pool test

---
 metagpt/exp_pool/schema.py             |  5 +-
 tests/metagpt/exp_pool/test_manager.py | 77 ++++++++++++++++++++++++++
 2 files changed, 80 insertions(+), 2 deletions(-)
 create mode 100644 tests/metagpt/exp_pool/test_manager.py

diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index e6ae4ee1d..1afcc1508 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -1,7 +1,7 @@
 """Experience schema."""
 
 from enum import Enum
-from typing import Optional
+from typing import Any, Optional
 
 from llama_index.core.schema import TextNode
 from pydantic import BaseModel, Field
@@ -38,13 +38,14 @@ class Trajectory(BaseModel):
     plan: str = Field(default="", description="The plan.")
     action: str = Field(default="", description="Action for the plan.")
     observation: str = Field(default="", description="Output of the action.")
+    reward: int = Field(default=0, description="Measure the action.")
 
 
 class Experience(BaseModel):
     """Experience."""
 
     req: str = Field(..., description="")
-    resp: str = Field(..., description="The type is string/json/code.")
+    resp: Any = Field(..., description="The type is string/json/code.")
     metric: Optional[Metric] = Field(default=None, description="Metric.")
     exp_type: ExperienceType = Field(default=ExperienceType.SUCCESS, description="The type of experience.")
     entry_type: EntryType = Field(default=EntryType.AUTOMATIC, description="Type of entry: Manual or Automatic.")
diff --git a/tests/metagpt/exp_pool/test_manager.py b/tests/metagpt/exp_pool/test_manager.py
new file mode 100644
index 000000000..a0d7005f5
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_manager.py
@@ -0,0 +1,77 @@
+import pytest
+
+from metagpt.config2 import Config
+from metagpt.configs.exp_pool_config import ExperiencePoolConfig
+from metagpt.configs.llm_config import LLMConfig
+from metagpt.exp_pool.manager import ExperienceManager
+from metagpt.exp_pool.schema import MAX_SCORE, Experience, Metric
+from metagpt.rag.engines import SimpleEngine
+
+
+class TestExperienceManager:
+    @pytest.fixture
+    def mock_config(self):
+        return Config(llm=LLMConfig(), exp_pool=ExperiencePoolConfig(enable_write=True, enable_read=True))
+
+    @pytest.fixture
+    def mock_storage(self, mocker):
+        engine = mocker.MagicMock(spec=SimpleEngine)
+        engine.add_objs = mocker.MagicMock()
+        engine.aretrieve = mocker.AsyncMock(return_value=[])
+        return engine
+
+    @pytest.fixture
+    def mock_experience_manager(self, mock_config, mock_storage):
+        return ExperienceManager(config=mock_config, storage=mock_storage)
+
+    @pytest.fixture
+    def mock_experience(self):
+        return Experience(req="req", resp="resp")
+
+    def test_initialize_storage(self, mock_experience_manager, mock_storage):
+        assert mock_experience_manager.storage is mock_storage
+
+    def test_create_exp(self, mock_experience_manager, mock_experience):
+        mock_experience_manager.create_exp(mock_experience)
+        mock_experience_manager.storage.add_objs.assert_called_once_with([mock_experience])
+
+    def test_create_exp_write_disabled(self, mock_experience_manager, mock_experience, mock_config):
+        mock_config.exp_pool.enable_write = False
+        mock_experience_manager.create_exp(mock_experience)
+        mock_experience_manager.storage.add_objs.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_query_exps(self, mock_experience_manager, mocker):
+        req = "req"
+        resp = "resp"
+        tag = "test"
+        experiences = [Experience(req=req, resp=resp, tag="test"), Experience(req=req, resp=resp, tag="other")]
+        mock_experience_manager.storage.aretrieve.return_value = [
+            mocker.MagicMock(metadata={"obj": exp}) for exp in experiences
+        ]
+
+        result = await mock_experience_manager.query_exps(req, tag)
+        assert len(result) == 1
+        assert result[0].tag == "test"
+
+    @pytest.mark.asyncio
+    async def test_query_exps_no_read_permission(self, mock_experience_manager, mock_config):
+        mock_config.exp_pool.enable_read = False
+        result = await mock_experience_manager.query_exps("query")
+        assert result == []
+
+    def test_extract_one_perfect_exp(self, mock_experience_manager):
+        experiences = [
+            Experience(req="req", resp="resp", metric=Metric(score=MAX_SCORE)),
+            Experience(req="req", resp="resp"),
+        ]
+        perfect_exp: Experience = mock_experience_manager.extract_one_perfect_exp(experiences)
+        assert perfect_exp is not None
+        assert perfect_exp.metric.score == MAX_SCORE
+
+    def test_is_perfect_exp(self):
+        exp = Experience(req="req", resp="resp", metric=Metric(score=MAX_SCORE))
+        assert ExperienceManager.is_perfect_exp(exp) == True
+
+        exp = Experience(req="req", resp="resp")
+        assert ExperienceManager.is_perfect_exp(exp) == False

From 1d8d85e9a50f02ddf4b394ff1589c77322ac0c19 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Wed, 5 Jun 2024 10:32:48 +0800
Subject: [PATCH 07/51] update comment

---
 metagpt/exp_pool/decorator.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 1d691b8f3..e073ee494 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -22,18 +22,21 @@ def exp_cache(_func: Optional[Callable[..., ReturnType]] = None):
     def decorator(func: Callable[..., ReturnType]) -> Callable[..., ReturnType]:
         @functools.wraps(func)
         async def get_or_create(args: Any, kwargs: Any, is_async: bool) -> ReturnType:
-            """Attempts to retrieve a cached experience or creates one if not found."""
+            """Attempts to retrieve a perfect experience or creates an experience if not found."""
 
+            # 1. Get exps.
             req = f"{func.__name__}_{args}_{kwargs}"
             exps = await exp_manager.query_exps(req)
             if perfect_exp := exp_manager.extract_one_perfect_exp(exps):
                 return perfect_exp
 
+            # 2. Exec func. TODO: pass exps to func
             if is_async:
                 result = await func(*args, **kwargs)
             else:
                 result = func(*args, **kwargs)
 
+            # 3. Create an exp.
             exp_manager.create_exp(Experience(req=req, resp=result))
 
             return result

From c78cddd1021c5073e7f4e17d22149053fc8e3276 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Wed, 5 Jun 2024 22:15:09 +0800
Subject: [PATCH 08/51] add exp_pool tests

---
 examples/exp_pool/decorator.py           |   5 +-
 metagpt/exp_pool/decorator.py            | 144 ++++++++++++++++------
 metagpt/exp_pool/manager.py              |  10 +-
 metagpt/exp_pool/schema.py               |  16 ++-
 metagpt/exp_pool/scorers/__init__.py     |   6 +
 metagpt/exp_pool/scorers/base.py         |  27 +++++
 metagpt/exp_pool/scorers/simple.py       |  73 ++++++++++++
 tests/metagpt/exp_pool/test_decorator.py | 145 +++++++++++++++++++++++
 tests/metagpt/exp_pool/test_manager.py   |   8 +-
 9 files changed, 391 insertions(+), 43 deletions(-)
 create mode 100644 metagpt/exp_pool/scorers/__init__.py
 create mode 100644 metagpt/exp_pool/scorers/base.py
 create mode 100644 metagpt/exp_pool/scorers/simple.py
 create mode 100644 tests/metagpt/exp_pool/test_decorator.py

diff --git a/examples/exp_pool/decorator.py b/examples/exp_pool/decorator.py
index 2f6397f80..3f6093e01 100644
--- a/examples/exp_pool/decorator.py
+++ b/examples/exp_pool/decorator.py
@@ -7,8 +7,9 @@ from metagpt.exp_pool import exp_cache, exp_manager
 from metagpt.logs import logger
 
 
-@exp_cache
-async def produce(req):
+@exp_cache(pass_exps_to_func=True)
+async def produce(req, exps=None):
+    logger.info(f"Previous experiences: {exps}")
     return f"{req} {uuid.uuid4().hex}"
 
 
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index e073ee494..9eb4d9e61 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -4,56 +4,134 @@ import asyncio
 import functools
 from typing import Any, Callable, Optional, TypeVar
 
-from metagpt.exp_pool.manager import exp_manager
-from metagpt.exp_pool.schema import Experience
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.exp_pool.manager import ExperienceManager, exp_manager
+from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
+from metagpt.exp_pool.scorers import ExperienceScorer, SimpleScorer
 from metagpt.utils.async_helper import NestAsyncio
 
 ReturnType = TypeVar("ReturnType")
 
 
-def exp_cache(_func: Optional[Callable[..., ReturnType]] = None):
-    """Decorator to check for a perfect experience and returns it if exists.
-
-    Otherwise, it executes the function, save the result as a new experience, and returns the result.
+def exp_cache(
+    _func: Optional[Callable[..., ReturnType]] = None,
+    query_type: QueryType = QueryType.SEMANTIC,
+    scorer: Optional[ExperienceScorer] = None,
+    manager: Optional[ExperienceManager] = None,
+    pass_exps_to_func: bool = False,
+):
+    """Decorator to get a perfect experience, otherwise, it executes the function, and create a new experience.
 
     This can be applied to both synchronous and asynchronous functions.
+
+    Args:
+        _func: Just to make the decorator more flexible, for example, it can be used directly with @exp_cache by default, without the need for @exp_cache().
+        query_type: The type of query to be used when fetching experiences.
+        scorer: Evaluate experience. Default SimpleScorer.
+        manager: How to fetch, evaluate and save experience, etc. Default exp_manager.
+        pass_exps_to_func: To control whether imperfect experiences are passed to the function, if True, the func must have a parameter named 'exps'.
     """
 
     def decorator(func: Callable[..., ReturnType]) -> Callable[..., ReturnType]:
         @functools.wraps(func)
-        async def get_or_create(args: Any, kwargs: Any, is_async: bool) -> ReturnType:
-            """Attempts to retrieve a perfect experience or creates an experience if not found."""
+        async def get_or_create(args: Any, kwargs: Any) -> ReturnType:
+            handler = ExpCacheHandler(
+                func=func,
+                args=args,
+                kwargs=kwargs,
+                exp_manager=manager or exp_manager,
+                exp_scorer=scorer or SimpleScorer(),
+                pass_exps=pass_exps_to_func,
+            )
 
-            # 1. Get exps.
-            req = f"{func.__name__}_{args}_{kwargs}"
-            exps = await exp_manager.query_exps(req)
-            if perfect_exp := exp_manager.extract_one_perfect_exp(exps):
-                return perfect_exp
+            await handler.fetch_experiences(query_type)
+            if exp := handler.get_one_perfect_experience():
+                return exp
 
-            # 2. Exec func. TODO: pass exps to func
-            if is_async:
-                result = await func(*args, **kwargs)
-            else:
-                result = func(*args, **kwargs)
+            await handler.execute_function()
+            await handler.evaluate_experience()
+            handler.save_experience()
 
-            # 3. Create an exp.
-            exp_manager.create_exp(Experience(req=req, resp=result))
+            return handler._result
 
-            return result
+        return ExpCacheHandler.choose_wrapper(func, get_or_create)
 
-        def sync_wrapper(*args: Any, **kwargs: Any) -> ReturnType:
+    return decorator(_func) if _func else decorator
+
+
+class ExpCacheHandler(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    func: Callable
+    args: Any
+    kwargs: Any
+    exp_manager: ExperienceManager
+    exp_scorer: ExperienceScorer
+    pass_exps: bool
+
+    _exps: list[Experience] = None
+    _result: Any = None
+    _score: Score = None
+
+    async def fetch_experiences(self, query_type: QueryType):
+        """Fetch a potentially perfect existing experience."""
+
+        req = self.generate_req_identifier()
+        self._exps = await self.exp_manager.query_exps(req, query_type=query_type)
+
+    def get_one_perfect_experience(self) -> Optional[Experience]:
+        return self.exp_manager.extract_one_perfect_exp(self._exps)
+
+    async def execute_function(self):
+        """Execute the function, and save the result."""
+        self._result = await self._execute_function()
+
+    async def evaluate_experience(self):
+        """Evaluate the experience, and save the score."""
+
+        self._score = await self.exp_scorer.evaluate(self.func, self._result, self.args, self.kwargs)
+
+    def save_experience(self):
+        """Save the new experience."""
+
+        req = self.generate_req_identifier()
+        exp = Experience(req=req, resp=self._result, metric=Metric(score=self._score))
+
+        self.exp_manager.create_exp(exp)
+
+    def generate_req_identifier(self):
+        """Generate a unique request identifier based on the function and its arguments."""
+
+        return f"{self.func.__name__}_{self.args}_{self.kwargs}"
+
+    @staticmethod
+    def choose_wrapper(func, wrapped_func):
+        """Choose how to run wrapped_func based on whether the function is asynchronous."""
+
+        async def async_wrapper(*args, **kwargs):
+            return await wrapped_func(args, kwargs)
+
+        def sync_wrapper(*args, **kwargs):
             NestAsyncio.apply_once()
-            return asyncio.get_event_loop().run_until_complete(get_or_create(args, kwargs, is_async=False))
+            return asyncio.get_event_loop().run_until_complete(wrapped_func(args, kwargs))
 
-        async def async_wrapper(*args: Any, **kwargs: Any) -> ReturnType:
-            return await get_or_create(args, kwargs, is_async=True)
+        return async_wrapper if asyncio.iscoroutinefunction(func) else sync_wrapper
 
-        if asyncio.iscoroutinefunction(func):
-            return async_wrapper
-        else:
-            return sync_wrapper
+    async def _execute_function(self):
+        if self.pass_exps:
+            return await self._execute_function_with_exps()
 
-    if _func is None:
-        return decorator
-    else:
-        return decorator(_func)
+        return await self._execute_function_without_exps()
+
+    async def _execute_function_without_exps(self):
+        if asyncio.iscoroutinefunction(self.func):
+            return await self.func(*self.args, **self.kwargs)
+
+        return self.func(*self.args, **self.kwargs)
+
+    async def _execute_function_with_exps(self):
+        if asyncio.iscoroutinefunction(self.func):
+            return await self.func(*self.args, **self.kwargs, exps=self._exps)
+
+        return self.func(*self.args, **self.kwargs, exps=self._exps)
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 4bc566104..58499104d 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -5,7 +5,7 @@ from typing import Optional
 from pydantic import BaseModel, ConfigDict, model_validator
 
 from metagpt.config2 import Config, config
-from metagpt.exp_pool.schema import MAX_SCORE, Experience
+from metagpt.exp_pool.schema import MAX_SCORE, Experience, QueryType
 from metagpt.rag.engines import SimpleEngine
 from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
 
@@ -45,12 +45,13 @@ class ExperienceManager(BaseModel):
 
         self.storage.add_objs([exp])
 
-    async def query_exps(self, req: str, tag: str = "") -> list[Experience]:
+    async def query_exps(self, req: str, tag: str = "", query_type: QueryType = QueryType.SEMANTIC) -> list[Experience]:
         """Retrieves and filters experiences.
 
         Args:
             req (str): The query string to retrieve experiences.
             tag (str): Optional tag to filter the experiences by.
+            query_type (QueryType): Default semantic to vector matching. exact to same matching.
 
         Returns:
             list[Experience]: A list of experiences that match the args.
@@ -65,6 +66,9 @@ class ExperienceManager(BaseModel):
         if tag:
             exps = [exp for exp in exps if exp.tag == tag]
 
+        if query_type == QueryType.EXACT:
+            exps = [exp for exp in exps if exp.req == req]
+
         return exps
 
     def extract_one_perfect_exp(self, exps: list[Experience]) -> Optional[Experience]:
@@ -96,7 +100,7 @@ class ExperienceManager(BaseModel):
             return False
 
         # TODO: need more metrics
-        if exp.metric and exp.metric.score == MAX_SCORE:
+        if exp.metric and exp.metric.score.val == MAX_SCORE:
             return True
 
         return False
diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index 1afcc1508..9fc665cca 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -9,6 +9,13 @@ from pydantic import BaseModel, Field
 MAX_SCORE = 10
 
 
+class QueryType(str, Enum):
+    """Type of query experiences."""
+
+    EXACT = "exact"
+    SEMANTIC = "semantic"
+
+
 class ExperienceType(str, Enum):
     """Experience Type."""
 
@@ -24,12 +31,19 @@ class EntryType(Enum):
     MANUAL = "Manual"
 
 
+class Score(BaseModel):
+    """Score in Metric."""
+
+    val: int = Field(default=1, description="Value of the score, Between 1 and 10, higher is better.")
+    reason: str = Field(default="", description="Reason for the value.")
+
+
 class Metric(BaseModel):
     """Experience Metric."""
 
     time_cost: float = Field(default=0.000, description="Time cost, the unit is milliseconds.")
     money_cost: float = Field(default=0.000, description="Money cost, the unit is US dollars.")
-    score: int = Field(default=1, description="Score, a value between 1 and 10.")
+    score: Score = Field(default=None, description="Score, with value and reason.")
 
 
 class Trajectory(BaseModel):
diff --git a/metagpt/exp_pool/scorers/__init__.py b/metagpt/exp_pool/scorers/__init__.py
new file mode 100644
index 000000000..85bea88ff
--- /dev/null
+++ b/metagpt/exp_pool/scorers/__init__.py
@@ -0,0 +1,6 @@
+"""Experience scorers init."""
+
+from metagpt.exp_pool.scorers.base import ExperienceScorer
+from metagpt.exp_pool.scorers.simple import SimpleScorer
+
+__all__ = ["ExperienceScorer", "SimpleScorer"]
diff --git a/metagpt/exp_pool/scorers/base.py b/metagpt/exp_pool/scorers/base.py
new file mode 100644
index 000000000..a9d30cffe
--- /dev/null
+++ b/metagpt/exp_pool/scorers/base.py
@@ -0,0 +1,27 @@
+"""Experience Scorers."""
+
+from abc import abstractmethod
+from typing import Any, Callable
+
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.exp_pool.schema import Score
+
+
+class ExperienceScorer(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    @abstractmethod
+    async def evaluate(self, func: Callable, result: Any, args: tuple = None, kwargs: dict = None) -> Score:
+        """Evaluate the quality of the result produced by the function and parameters.
+
+        Args:
+            func (Callable): The function whose result is to be evaluated.
+            result (Any): The result produced by the function.
+            args (Tuple[Any, ...]): The tuple of arguments that were passed to the function.
+            kwargs (Dict[str, Any]): The dictionary of keyword arguments that were passed to the function.
+
+        Example:
+            result = await sample(5, name="foo")
+            score = await scorer.evaluate(sample, result, args=(5), kwargs={"name": "foo"})
+        """
diff --git a/metagpt/exp_pool/scorers/simple.py b/metagpt/exp_pool/scorers/simple.py
new file mode 100644
index 000000000..d0301cbc2
--- /dev/null
+++ b/metagpt/exp_pool/scorers/simple.py
@@ -0,0 +1,73 @@
+"""Evalate by llm."""
+import inspect
+import json
+from typing import Any, Callable
+
+from pydantic import Field
+
+from metagpt.exp_pool.schema import Score
+from metagpt.exp_pool.scorers.base import ExperienceScorer
+from metagpt.llm import LLM
+from metagpt.provider.base_llm import BaseLLM
+from metagpt.utils.common import parse_json_code_block
+
+SIMPLE_SCORER_TEMPLATE = """
+Role: You're an expert score evaluator. You specialize in assessing the output of the given function, based on its intended requirement and produced result.
+
+## Context
+### Function Name
+{func_name}
+
+### Function Document
+{func_doc}
+
+### Function Signature
+{func_signature}
+
+### Function Parameters
+args: {func_args}
+kwargs: {func_kwargs}
+
+### Produced Result By Function and Parameters
+{func_result}
+
+## Format Example
+```json
+{{
+    "val": "the value of the score, int from 1 to 10, higher is better.",
+    "reason": "an explanation supporting the score."
+}}
+```
+
+## Instructions
+- Understand the function and requirements given by the user.
+- Analyze the results produced by the function.
+- Grade the results based on level of alignment with the requirements.
+- Provide a score on a scale defined by user or a default scale (1 to 10).
+
+## Constraint
+Format: Just print the result in json format like **Format Example**.
+
+## Action
+Follow instructions, generate output and make sure it follows the **Constraint**.
+"""
+
+
+class SimpleScorer(ExperienceScorer):
+    llm: BaseLLM = Field(default_factory=LLM)
+
+    async def evaluate(self, func: Callable, result: Any, args: tuple = None, kwargs: dict = None) -> Score:
+        """Evaluate the quality of content."""
+
+        prompt = SIMPLE_SCORER_TEMPLATE.format(
+            func_name=func.__name__,
+            func_doc=func.__doc__,
+            func_signature=inspect.signature(func),
+            func_args=args,
+            func_kwargs=kwargs,
+            func_result=result,
+        )
+        resp = await self.llm.aask(prompt)
+        resp_json = json.loads(parse_json_code_block(resp)[0])
+
+        return Score(**resp_json)
diff --git a/tests/metagpt/exp_pool/test_decorator.py b/tests/metagpt/exp_pool/test_decorator.py
new file mode 100644
index 000000000..508229d18
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_decorator.py
@@ -0,0 +1,145 @@
+import asyncio
+
+import pytest
+
+from metagpt.exp_pool.decorator import ExpCacheHandler
+from metagpt.exp_pool.manager import ExperienceManager
+from metagpt.exp_pool.schema import Experience, QueryType, Score
+from metagpt.exp_pool.scorers import SimpleScorer
+from metagpt.rag.engines import SimpleEngine
+
+
+class TestExpCache:
+    @pytest.fixture
+    def mock_func(self, mocker):
+        return mocker.AsyncMock()
+
+    @pytest.fixture
+    def mock_exp_manager(self, mocker):
+        manager = mocker.MagicMock(spec=ExperienceManager)
+        manager.storage = mocker.MagicMock(spec=SimpleEngine)
+        manager.query_exps = mocker.AsyncMock()
+        manager.create_exp = mocker.MagicMock()
+        manager.extract_one_perfect_exp = mocker.MagicMock()
+        return manager
+
+    @pytest.fixture
+    def mock_scorer(self, mocker):
+        scorer = mocker.MagicMock(spec=SimpleScorer)
+        scorer.evaluate = mocker.AsyncMock()
+        return scorer
+
+    @pytest.fixture
+    def exp_cache_handler(self, mock_func, mock_exp_manager, mock_scorer):
+        return ExpCacheHandler(
+            func=mock_func, args=(), kwargs={}, exp_manager=mock_exp_manager, exp_scorer=mock_scorer, pass_exps=False
+        )
+
+    @pytest.mark.asyncio
+    async def test_fetch_experiences(self, exp_cache_handler, mock_exp_manager):
+        await exp_cache_handler.fetch_experiences(QueryType.SEMANTIC)
+        mock_exp_manager.query_exps.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_perfect_experience_found(self, exp_cache_handler, mock_exp_manager, mock_func):
+        # Setup: Assume perfect experience is found
+        perfect_exp = Experience(req="req", resp="resp")
+        mock_exp_manager.extract_one_perfect_exp.return_value = perfect_exp
+
+        # Execute
+        exp_cache_handler._exps = [perfect_exp]  # Simulate fetched experiences
+        result = exp_cache_handler.get_one_perfect_experience()
+
+        # Assert
+        assert result.resp == "resp"
+        mock_func.assert_not_called()  # Function should not be called
+
+    @pytest.mark.asyncio
+    async def test_execute_function_when_no_perfect_exp(self, exp_cache_handler, mock_exp_manager, mock_func):
+        # Setup: No perfect experience
+        mock_exp_manager.extract_one_perfect_exp.return_value = None
+        mock_func.return_value = "Computed result"
+
+        # Execute
+        await exp_cache_handler.execute_function()
+
+        # Assert
+        assert exp_cache_handler._result == "Computed result"
+        mock_func.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_evaluate_and_save_experience(self, exp_cache_handler, mock_scorer, mock_exp_manager):
+        # Setup
+        mock_scorer.evaluate.return_value = Score(value=100)
+        exp_cache_handler._result = "Computed result"
+
+        # Execute
+        await exp_cache_handler.evaluate_experience()
+        exp_cache_handler.save_experience()
+
+        # Assert
+        mock_scorer.evaluate.assert_called_once()
+        mock_exp_manager.create_exp.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_async_function_execution_with_exps(self, exp_cache_handler, mock_exp_manager, mock_func):
+        # Setup
+        exp_cache_handler.pass_exps = True
+        mock_func.return_value = "Async result with exps"
+        mock_exp_manager.extract_one_perfect_exp.return_value = None
+        exp_cache_handler._exps = [Experience(req="req", resp="resp")]
+
+        # Execute
+        await exp_cache_handler.execute_function()
+
+        # Assert
+        mock_func.assert_called_once_with(exps=exp_cache_handler._exps)
+        assert exp_cache_handler._result == "Async result with exps"
+
+    def test_sync_function_execution_with_exps(self, mocker, exp_cache_handler, mock_exp_manager, mock_func):
+        # Setup
+        exp_cache_handler.func = mocker.Mock(return_value="Sync result with exps")
+        exp_cache_handler.pass_exps = True
+        mock_exp_manager.extract_one_perfect_exp.return_value = None
+        exp_cache_handler._exps = [Experience(req="req", resp="resp")]
+
+        # Execute
+        asyncio.get_event_loop().run_until_complete(exp_cache_handler.execute_function())
+
+        # Assert
+        exp_cache_handler.func.assert_called_once_with(exps=exp_cache_handler._exps)
+        assert exp_cache_handler._result == "Sync result with exps"
+
+    def test_wrapper_selection_async(self, mocker, exp_cache_handler, mock_func):
+        # Setup
+        mock_func = mocker.AsyncMock()
+
+        # Execute
+        wrapper = ExpCacheHandler.choose_wrapper(mock_func, exp_cache_handler.execute_function)
+
+        # Assert
+        assert asyncio.iscoroutinefunction(wrapper), "Wrapper should be asynchronous"
+
+    def test_wrapper_selection_sync(self, exp_cache_handler, mocker):
+        # Setup
+        sync_func = mocker.Mock()
+
+        # Execute
+        wrapper = ExpCacheHandler.choose_wrapper(sync_func, exp_cache_handler.execute_function)
+
+        # Assert
+        assert not asyncio.iscoroutinefunction(wrapper), "Wrapper should be synchronous"
+
+    @pytest.mark.asyncio
+    async def test_generate_req_identifier(self, exp_cache_handler):
+        # Setup
+        exp_cache_handler.func = lambda x: x
+        exp_cache_handler.args = (42,)
+        exp_cache_handler.kwargs = {"y": 3.14}
+
+        # Execute
+        req_id = exp_cache_handler.generate_req_identifier()
+
+        # Assert
+        expected_id = "<lambda>_(42,)_{'y': 3.14}"
+        assert req_id == expected_id, "Request identifier should match the expected format"
diff --git a/tests/metagpt/exp_pool/test_manager.py b/tests/metagpt/exp_pool/test_manager.py
index a0d7005f5..3e8f47417 100644
--- a/tests/metagpt/exp_pool/test_manager.py
+++ b/tests/metagpt/exp_pool/test_manager.py
@@ -4,7 +4,7 @@ from metagpt.config2 import Config
 from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 from metagpt.configs.llm_config import LLMConfig
 from metagpt.exp_pool.manager import ExperienceManager
-from metagpt.exp_pool.schema import MAX_SCORE, Experience, Metric
+from metagpt.exp_pool.schema import MAX_SCORE, Experience, Metric, Score
 from metagpt.rag.engines import SimpleEngine
 
 
@@ -62,15 +62,15 @@ class TestExperienceManager:
 
     def test_extract_one_perfect_exp(self, mock_experience_manager):
         experiences = [
-            Experience(req="req", resp="resp", metric=Metric(score=MAX_SCORE)),
+            Experience(req="req", resp="resp", metric=Metric(score=Score(val=MAX_SCORE))),
             Experience(req="req", resp="resp"),
         ]
         perfect_exp: Experience = mock_experience_manager.extract_one_perfect_exp(experiences)
         assert perfect_exp is not None
-        assert perfect_exp.metric.score == MAX_SCORE
+        assert perfect_exp.metric.score.val == MAX_SCORE
 
     def test_is_perfect_exp(self):
-        exp = Experience(req="req", resp="resp", metric=Metric(score=MAX_SCORE))
+        exp = Experience(req="req", resp="resp", metric=Metric(score=Score(val=MAX_SCORE)))
         assert ExperienceManager.is_perfect_exp(exp) == True
 
         exp = Experience(req="req", resp="resp")

From d148a3217bbe6eb2aa80bdd0132801da98fd1a14 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Wed, 5 Jun 2024 23:26:09 +0800
Subject: [PATCH 09/51] add handle_exception to ensure robustness

---
 metagpt/exp_pool/decorator.py | 14 ++++++++++++--
 metagpt/exp_pool/manager.py   |  3 +++
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 9eb4d9e61..9cf924779 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -10,6 +10,7 @@ from metagpt.exp_pool.manager import ExperienceManager, exp_manager
 from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
 from metagpt.exp_pool.scorers import ExperienceScorer, SimpleScorer
 from metagpt.utils.async_helper import NestAsyncio
+from metagpt.utils.exceptions import handle_exception
 
 ReturnType = TypeVar("ReturnType")
 
@@ -50,8 +51,7 @@ def exp_cache(
                 return exp
 
             await handler.execute_function()
-            await handler.evaluate_experience()
-            handler.save_experience()
+            await handler.process_experience()
 
             return handler._result
 
@@ -87,6 +87,16 @@ class ExpCacheHandler(BaseModel):
         """Execute the function, and save the result."""
         self._result = await self._execute_function()
 
+    @handle_exception
+    async def process_experience(self):
+        """Process experience.
+
+        Evaluates and saves experience.
+        Use `handle_exception` to ensure robustness, do not stop subsequent operations.
+        """
+        await self.evaluate_experience()
+        self.save_experience()
+
     async def evaluate_experience(self):
         """Evaluate the experience, and save the score."""
 
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 58499104d..546086b37 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -8,6 +8,7 @@ from metagpt.config2 import Config, config
 from metagpt.exp_pool.schema import MAX_SCORE, Experience, QueryType
 from metagpt.rag.engines import SimpleEngine
 from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
+from metagpt.utils.exceptions import handle_exception
 
 
 class ExperienceManager(BaseModel):
@@ -34,6 +35,7 @@ class ExperienceManager(BaseModel):
             )
         return self
 
+    @handle_exception
     def create_exp(self, exp: Experience):
         """Adds an experience to the storage if writing is enabled.
 
@@ -45,6 +47,7 @@ class ExperienceManager(BaseModel):
 
         self.storage.add_objs([exp])
 
+    @handle_exception(default_return=[])
     async def query_exps(self, req: str, tag: str = "", query_type: QueryType = QueryType.SEMANTIC) -> list[Experience]:
         """Retrieves and filters experiences.
 

From 1679757d9f3cc086ab251d0c6be4e542f1f14830 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Thu, 6 Jun 2024 20:18:40 +0800
Subject: [PATCH 10/51] update exp_pool example

---
 examples/exp_pool/simple.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/simple.py
index f270824bf..3216e78b8 100644
--- a/examples/exp_pool/simple.py
+++ b/examples/exp_pool/simple.py
@@ -9,8 +9,7 @@ from metagpt.logs import logger
 
 async def main():
     req = "Simple task."
-    resp = "Simple echo."
-    exp = Experience(req=req, resp=resp, entry_type=EntryType.MANUAL)
+    exp = Experience(req=req, resp="echo", entry_type=EntryType.MANUAL)
 
     exp_manager.create_exp(exp)
     logger.info(f"New experience created for the request `{req}`.")

From 16fd197e068cd10947340382d246ab505d5f0860 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Fri, 7 Jun 2024 10:30:37 +0800
Subject: [PATCH 11/51] update comment

---
 metagpt/exp_pool/scorers/simple.py | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/metagpt/exp_pool/scorers/simple.py b/metagpt/exp_pool/scorers/simple.py
index d0301cbc2..5779f7fb1 100644
--- a/metagpt/exp_pool/scorers/simple.py
+++ b/metagpt/exp_pool/scorers/simple.py
@@ -1,4 +1,5 @@
-"""Evalate by llm."""
+"""Simple Scorer."""
+
 import inspect
 import json
 from typing import Any, Callable
@@ -57,8 +58,17 @@ class SimpleScorer(ExperienceScorer):
     llm: BaseLLM = Field(default_factory=LLM)
 
     async def evaluate(self, func: Callable, result: Any, args: tuple = None, kwargs: dict = None) -> Score:
-        """Evaluate the quality of content."""
+        """Evaluates the quality of content by LLM.
 
+        Args:
+            func: The function to evaluate.
+            result: The result produced by the function.
+            args: The positional arguments used when calling the function, if any.
+            kwargs: The keyword arguments used when calling the function, if any.
+
+        Returns:
+            A Score object containing the evaluation results.
+        """
         prompt = SIMPLE_SCORER_TEMPLATE.format(
             func_name=func.__name__,
             func_doc=func.__doc__,

From 547bbfcffc2086ea3958fbf829aab55ec9e3e17d Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Fri, 7 Jun 2024 14:35:47 +0800
Subject: [PATCH 12/51] update comment

---
 metagpt/exp_pool/manager.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 546086b37..35ee5fdac 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -14,7 +14,7 @@ from metagpt.utils.exceptions import handle_exception
 class ExperienceManager(BaseModel):
     """ExperienceManager manages the lifecycle of experiences, including CRUD and optimization.
 
-    Attributes:
+    Args:
         config (Config): Configuration for managing experiences.
         storage (SimpleEngine): Engine to handle the storage and retrieval of experiences.
     """

From 7ac8397cc97ce7ad361f9420715ae55b5c1b5d88 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Fri, 7 Jun 2024 18:15:23 +0800
Subject: [PATCH 13/51] add scorer example

---
 examples/exp_pool/{simple.py => manager.py} |  0
 examples/exp_pool/scorer.py                 | 25 +++++++++++++++++++++
 metagpt/exp_pool/scorers/simple.py          |  4 ++--
 3 files changed, 27 insertions(+), 2 deletions(-)
 rename examples/exp_pool/{simple.py => manager.py} (100%)
 create mode 100644 examples/exp_pool/scorer.py

diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/manager.py
similarity index 100%
rename from examples/exp_pool/simple.py
rename to examples/exp_pool/manager.py
diff --git a/examples/exp_pool/scorer.py b/examples/exp_pool/scorer.py
new file mode 100644
index 000000000..1efe07bdf
--- /dev/null
+++ b/examples/exp_pool/scorer.py
@@ -0,0 +1,25 @@
+import asyncio
+
+from metagpt.exp_pool.scorers import SimpleScorer
+from metagpt.logs import logger
+
+
+def echo(req: str):
+    """Echo from req."""
+
+    return req
+
+
+async def simple():
+    scorer = SimpleScorer()
+
+    score = await scorer.evaluate(echo, "data", ("data",))
+    logger.info(f"The score is: {score}")
+
+
+async def main():
+    await simple()
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/metagpt/exp_pool/scorers/simple.py b/metagpt/exp_pool/scorers/simple.py
index 5779f7fb1..84995b60f 100644
--- a/metagpt/exp_pool/scorers/simple.py
+++ b/metagpt/exp_pool/scorers/simple.py
@@ -10,7 +10,7 @@ from metagpt.exp_pool.schema import Score
 from metagpt.exp_pool.scorers.base import ExperienceScorer
 from metagpt.llm import LLM
 from metagpt.provider.base_llm import BaseLLM
-from metagpt.utils.common import parse_json_code_block
+from metagpt.utils.common import CodeParser
 
 SIMPLE_SCORER_TEMPLATE = """
 Role: You're an expert score evaluator. You specialize in assessing the output of the given function, based on its intended requirement and produced result.
@@ -78,6 +78,6 @@ class SimpleScorer(ExperienceScorer):
             func_result=result,
         )
         resp = await self.llm.aask(prompt)
-        resp_json = json.loads(parse_json_code_block(resp)[0])
+        resp_json = json.loads(CodeParser.parse_code(resp, lang="json"))
 
         return Score(**resp_json)

From 797a8c5326c13feebaa1d0676b7eebf571eb980e Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 11 Jun 2024 15:40:01 +0800
Subject: [PATCH 14/51] change req in exp

---
 metagpt/exp_pool/decorator.py          | 29 +++++++++++++++++++-------
 metagpt/exp_pool/manager.py            |  3 +--
 metagpt/utils/reflection.py            | 25 +++++++++++++++++-----
 tests/metagpt/utils/test_reflection.py | 29 ++++++++++++++++++++++++++
 4 files changed, 71 insertions(+), 15 deletions(-)
 create mode 100644 tests/metagpt/utils/test_reflection.py

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 9cf924779..e559797a3 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -11,6 +11,7 @@ from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
 from metagpt.exp_pool.scorers import ExperienceScorer, SimpleScorer
 from metagpt.utils.async_helper import NestAsyncio
 from metagpt.utils.exceptions import handle_exception
+from metagpt.utils.reflection import get_class_name
 
 ReturnType = TypeVar("ReturnType")
 
@@ -43,7 +44,7 @@ def exp_cache(
                 kwargs=kwargs,
                 exp_manager=manager or exp_manager,
                 exp_scorer=scorer or SimpleScorer(),
-                pass_exps=pass_exps_to_func,
+                pass_exps_to_func=pass_exps_to_func,
             )
 
             await handler.fetch_experiences(query_type)
@@ -68,16 +69,17 @@ class ExpCacheHandler(BaseModel):
     kwargs: Any
     exp_manager: ExperienceManager
     exp_scorer: ExperienceScorer
-    pass_exps: bool
+    pass_exps_to_func: bool = False
 
     _exps: list[Experience] = None
     _result: Any = None
     _score: Score = None
+    _req: str = None
 
     async def fetch_experiences(self, query_type: QueryType):
         """Fetch a potentially perfect existing experience."""
 
-        req = self.generate_req_identifier()
+        req = self._get_req_identifier()
         self._exps = await self.exp_manager.query_exps(req, query_type=query_type)
 
     def get_one_perfect_experience(self) -> Optional[Experience]:
@@ -105,15 +107,26 @@ class ExpCacheHandler(BaseModel):
     def save_experience(self):
         """Save the new experience."""
 
-        req = self.generate_req_identifier()
+        req = self._get_req_identifier()
         exp = Experience(req=req, resp=self._result, metric=Metric(score=self._score))
 
         self.exp_manager.create_exp(exp)
 
-    def generate_req_identifier(self):
-        """Generate a unique request identifier based on the function and its arguments."""
+    def _get_req_identifier(self):
+        """Generate a unique request identifier based on the function and its arguments.
 
-        return f"{self.func.__name__}_{self.args}_{self.kwargs}"
+        Result Example:
+        - "write_prd-('2048',)-{}"
+        - "WritePRD.run-('2048',)-{}"
+        """
+        if not self._req:
+            cls_name = get_class_name(self.func, *self.args)
+            func_name = f"{cls_name}.{self.func.__name__}" if cls_name else self.func.__name__
+            args = self.args[1:] if cls_name and len(self.args) >= 1 else self.args
+
+            self._req = f"{func_name}-{args}-{self.kwargs}"
+
+        return self._req
 
     @staticmethod
     def choose_wrapper(func, wrapped_func):
@@ -129,7 +142,7 @@ class ExpCacheHandler(BaseModel):
         return async_wrapper if asyncio.iscoroutinefunction(func) else sync_wrapper
 
     async def _execute_function(self):
-        if self.pass_exps:
+        if self.pass_exps_to_func:
             return await self._execute_function_with_exps()
 
         return await self._execute_function_without_exps()
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 35ee5fdac..7382fe8f1 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -7,7 +7,7 @@ from pydantic import BaseModel, ConfigDict, model_validator
 from metagpt.config2 import Config, config
 from metagpt.exp_pool.schema import MAX_SCORE, Experience, QueryType
 from metagpt.rag.engines import SimpleEngine
-from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
+from metagpt.rag.schema import ChromaRetrieverConfig
 from metagpt.utils.exceptions import handle_exception
 
 
@@ -31,7 +31,6 @@ class ExperienceManager(BaseModel):
                 retriever_configs=[
                     ChromaRetrieverConfig(collection_name="experience_pool", persist_path=".chroma_exp_data")
                 ],
-                ranker_configs=[LLMRankerConfig()],
             )
         return self
 
diff --git a/metagpt/utils/reflection.py b/metagpt/utils/reflection.py
index 2683e5657..9b10a4b3e 100644
--- a/metagpt/utils/reflection.py
+++ b/metagpt/utils/reflection.py
@@ -19,9 +19,24 @@ def check_methods(C, *methods):
     return True
 
 
-def get_func_full_name(func, *args) -> str:
-    if inspect.ismethod(func) or (inspect.isfunction(func) and "self" in inspect.signature(func).parameters):
-        cls_name = args[0].__class__.__name__
-        return f"{func.__module__}.{cls_name}.{func.__name__}"
+def get_class_name(func, *args) -> str:
+    """Returns the class name of the object that a method belongs to.
 
-    return f"{func.__module__}.{func.__name__}"
+    - If `func` is a bound method, extracts the class name directly from the method.
+    - If `func` is an unbound method and `args` are provided, assumes the first argument is `self` and extracts the class name.
+    - Returns an empty string if neither condition is met.
+    """
+    if inspect.ismethod(func):
+        return func.__self__.__class__.__name__
+
+    if inspect.isfunction(func) and "self" in inspect.signature(func).parameters and args:
+        return args[0].__class__.__name__
+
+    return ""
+
+
+def get_func_or_method_name(func, *args) -> str:
+    """Function name, or method name with class name."""
+    cls_name = get_class_name(func, *args)
+
+    return f"{cls_name}.{func.__name__}" if cls_name else f"{func.__name__}"
diff --git a/tests/metagpt/utils/test_reflection.py b/tests/metagpt/utils/test_reflection.py
new file mode 100644
index 000000000..e78e1b400
--- /dev/null
+++ b/tests/metagpt/utils/test_reflection.py
@@ -0,0 +1,29 @@
+from metagpt.utils.reflection import get_func_or_method_name
+
+
+def simple_function():
+    pass
+
+
+class SampleClass:
+    def method(self):
+        pass
+
+
+class TestFunctionOrMethodName:
+    def test_simple_function(self):
+        assert get_func_or_method_name(simple_function) == "simple_function"
+
+    def test_class_method_without_args(self):
+        sample_instance = SampleClass()
+        assert get_func_or_method_name(sample_instance.method) == "SampleClass.method"
+
+    def test_class_method_with_args(self):
+        sample_instance = SampleClass()
+        assert get_func_or_method_name(SampleClass.method, sample_instance) == "SampleClass.method"
+
+    def test_function_with_no_args(self):
+        assert get_func_or_method_name(simple_function) == "simple_function"
+
+    def test_method_without_instance(self):
+        assert get_func_or_method_name(SampleClass.method) == "method"

From 29c61a7fa407ccad69e871a147a5b9ad22e958b8 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 3 Jun 2024 10:14:57 +0800
Subject: [PATCH 15/51] experience pool

---
 config/config2.example.yaml        |  3 +++
 examples/exp_pool/manager.py       | 21 ++++++++++++++++++++
 metagpt/config2.py                 |  4 ++++
 metagpt/configs/exp_pool_config.py |  6 ++++++
 metagpt/exp_pool/__init__.py       |  0
 metagpt/exp_pool/decorator.py      |  4 ++++
 metagpt/exp_pool/manager.py        | 32 ++++++++++++++++++++++++++++++
 metagpt/exp_pool/schema.py         | 25 +++++++++++++++++++++++
 metagpt/utils/reflection.py        |  9 +++++++++
 metagpt/utils/token_counter.py     |  2 ++
 10 files changed, 106 insertions(+)
 create mode 100644 examples/exp_pool/manager.py
 create mode 100644 metagpt/configs/exp_pool_config.py
 create mode 100644 metagpt/exp_pool/__init__.py
 create mode 100644 metagpt/exp_pool/decorator.py
 create mode 100644 metagpt/exp_pool/manager.py
 create mode 100644 metagpt/exp_pool/schema.py

diff --git a/config/config2.example.yaml b/config/config2.example.yaml
index f1158775b..c5ca6e767 100644
--- a/config/config2.example.yaml
+++ b/config/config2.example.yaml
@@ -74,6 +74,9 @@ s3:
   secure: false
   bucket: "test"
 
+experience_pool:
+  enable_read: false
+  enable_write: false
 
 azure_tts_subscription_key: "YOUR_SUBSCRIPTION_KEY"
 azure_tts_region: "eastus"
diff --git a/examples/exp_pool/manager.py b/examples/exp_pool/manager.py
new file mode 100644
index 000000000..f5766f9a5
--- /dev/null
+++ b/examples/exp_pool/manager.py
@@ -0,0 +1,21 @@
+from metagpt.exp_pool.manager import ExperiencePoolManager
+from metagpt.exp_pool.schema import Experience
+from pprint import pprint
+import asyncio
+# import logging
+# logging.basicConfig(level=logging.DEBUG)
+
+async def main():
+    req = "2048 game"
+    exp = Experience(req=req, resp="python code")
+
+    manager = ExperiencePoolManager()
+
+    # pprint(manager.storage.get())
+    # manager.create_exp(exp)
+    result = await manager.query_exp(req)
+    print(result)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
\ No newline at end of file
diff --git a/metagpt/config2.py b/metagpt/config2.py
index 717fe63a9..6f5a1add6 100644
--- a/metagpt/config2.py
+++ b/metagpt/config2.py
@@ -22,6 +22,7 @@ from metagpt.configs.search_config import SearchConfig
 from metagpt.configs.workspace_config import WorkspaceConfig
 from metagpt.const import CONFIG_ROOT, METAGPT_ROOT
 from metagpt.utils.yaml_model import YamlModel
+from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 
 
 class CLIParams(BaseModel):
@@ -71,6 +72,9 @@ class Config(CLIParams, YamlModel):
     enable_longterm_memory: bool = False
     code_review_k_times: int = 2
 
+    # Experience Pool Parameters
+    experience_pool: Optional[ExperiencePoolConfig] = None
+
     # Will be removed in the future
     metagpt_tti_url: str = ""
     language: str = "English"
diff --git a/metagpt/configs/exp_pool_config.py b/metagpt/configs/exp_pool_config.py
new file mode 100644
index 000000000..f7312d2de
--- /dev/null
+++ b/metagpt/configs/exp_pool_config.py
@@ -0,0 +1,6 @@
+from metagpt.utils.yaml_model import YamlModel
+
+
+class ExperiencePoolConfig(YamlModel):
+    enable_read: bool = False
+    enable_write: bool = False
diff --git a/metagpt/exp_pool/__init__.py b/metagpt/exp_pool/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
new file mode 100644
index 000000000..6629e8377
--- /dev/null
+++ b/metagpt/exp_pool/decorator.py
@@ -0,0 +1,4 @@
+
+
+def exp_cache(func):
+    pass
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
new file mode 100644
index 000000000..c32073a9f
--- /dev/null
+++ b/metagpt/exp_pool/manager.py
@@ -0,0 +1,32 @@
+from pydantic import BaseModel, ConfigDict
+from metagpt.exp_pool.schema import Experience
+import uuid
+import chromadb
+from chromadb import Collection, QueryResult
+from typing import Optional
+from metagpt.rag.engines import SimpleEngine
+from metagpt.rag.schema import ChromaRetrieverConfig
+
+
+class ExperiencePoolManager(BaseModel):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self._storage = None
+
+    @property
+    def storage(self) -> SimpleEngine:
+        if self._storage is None:
+            self._storage = SimpleEngine.from_objs(retriever_configs=[ChromaRetrieverConfig(collection_name="experience_pool", persist_path="./chroma_data")])
+        return self._storage
+    
+    def create_exp(self, exp: Experience):
+        self.storage.add_objs([exp])
+    
+    async def query_exp(self, req: str) -> list[Experience]:
+        nodes = await self.storage.aretrieve(req)
+        exps = [node.metadata["obj"] for node in nodes]
+
+        return exps
+
+    
+
diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
new file mode 100644
index 000000000..359268612
--- /dev/null
+++ b/metagpt/exp_pool/schema.py
@@ -0,0 +1,25 @@
+from pydantic import BaseModel, Field
+from llama_index.core.schema import TextNode
+
+
+class Experience(BaseModel):
+    req: str = Field(..., description="")
+    resp: str = Field(..., description="")
+
+    def rag_key(self):
+        return self.req
+
+
+class ExperienceNodeMetadata(BaseModel):
+    """Metadata of ExperienceNode."""
+
+    resp: str = Field(..., description="")
+
+
+class ExperienceNode(TextNode):
+    """ExperienceNode for RAG."""
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.excluded_llm_metadata_keys = list(ExperienceNodeMetadata.model_fields.keys())
+        self.excluded_embed_metadata_keys = self.excluded_llm_metadata_keys
diff --git a/metagpt/utils/reflection.py b/metagpt/utils/reflection.py
index 8b8237ae7..688831f06 100644
--- a/metagpt/utils/reflection.py
+++ b/metagpt/utils/reflection.py
@@ -1,4 +1,5 @@
 """class tools, including method inspection, class attributes, inheritance relationships, etc."""
+import inspect
 
 
 def check_methods(C, *methods):
@@ -16,3 +17,11 @@ def check_methods(C, *methods):
         else:
             return NotImplemented
     return True
+
+
+def get_func_full_name(func, *args) -> str:
+    if inspect.ismethod(func) or (inspect.isfunction(func) and "self" in inspect.signature(func).parameters):
+        cls_name = args[0].__class__.__name__
+        return f"{func.__module__}.{cls_name}.{func.__name__}"
+    
+    return f"{func.__module__}.{func.__name__}"
diff --git a/metagpt/utils/token_counter.py b/metagpt/utils/token_counter.py
index 0ba2daa89..496842a2d 100644
--- a/metagpt/utils/token_counter.py
+++ b/metagpt/utils/token_counter.py
@@ -150,6 +150,8 @@ TOKEN_MAX = {
     "gpt-4-1106-preview": 128000,
     "gpt-4-vision-preview": 128000,
     "gpt-4-1106-vision-preview": 128000,
+    "gpt-4-turbo": 128000,
+    "gpt-4o": 128000,
     "gpt-4": 8192,
     "gpt-4-0613": 8192,
     "gpt-4-32k": 32768,

From 808d65b4c3ec5198ced79f487422109468494406 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 10:28:39 +0800
Subject: [PATCH 16/51] experiment pool init

---
 .gitignore                         |   1 +
 config/config2.example.yaml        |   6 +-
 examples/exp_pool/decorator.py     |  26 +++++++
 examples/exp_pool/manager.py       |  21 ------
 examples/exp_pool/simple.py        |  29 ++++++++
 metagpt/config2.py                 |   4 +-
 metagpt/configs/exp_pool_config.py |   6 +-
 metagpt/exp_pool/__init__.py       |   6 ++
 metagpt/exp_pool/decorator.py      |  56 +++++++++++++-
 metagpt/exp_pool/manager.py        | 113 ++++++++++++++++++++++++-----
 metagpt/exp_pool/schema.py         |  40 +++++++++-
 metagpt/utils/reflection.py        |   2 +-
 12 files changed, 257 insertions(+), 53 deletions(-)
 create mode 100644 examples/exp_pool/decorator.py
 delete mode 100644 examples/exp_pool/manager.py
 create mode 100644 examples/exp_pool/simple.py

diff --git a/.gitignore b/.gitignore
index aa5edd74a..7c64829ad 100644
--- a/.gitignore
+++ b/.gitignore
@@ -162,6 +162,7 @@ examples/graph_store.json
 examples/image__vector_store.json
 examples/index_store.json
 .chroma
+.chroma_exp_data
 *~$*
 workspace/*
 tmp
diff --git a/config/config2.example.yaml b/config/config2.example.yaml
index c5ca6e767..c7b2cae2c 100644
--- a/config/config2.example.yaml
+++ b/config/config2.example.yaml
@@ -74,9 +74,9 @@ s3:
   secure: false
   bucket: "test"
 
-experience_pool:
-  enable_read: false
-  enable_write: false
+exp_pool:
+  enable_read: true
+  enable_write: true
 
 azure_tts_subscription_key: "YOUR_SUBSCRIPTION_KEY"
 azure_tts_region: "eastus"
diff --git a/examples/exp_pool/decorator.py b/examples/exp_pool/decorator.py
new file mode 100644
index 000000000..2f6397f80
--- /dev/null
+++ b/examples/exp_pool/decorator.py
@@ -0,0 +1,26 @@
+"""Decorator example of experience pool."""
+
+import asyncio
+import uuid
+
+from metagpt.exp_pool import exp_cache, exp_manager
+from metagpt.logs import logger
+
+
+@exp_cache
+async def produce(req):
+    return f"{req} {uuid.uuid4().hex}"
+
+
+async def main():
+    req = "Water"
+
+    resp = await produce(req)
+    logger.info(f"The resp of `produce{req}` is: {resp}")
+
+    exps = await exp_manager.query_exps(req)
+    logger.info(f"Find experiences: {exps}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/examples/exp_pool/manager.py b/examples/exp_pool/manager.py
deleted file mode 100644
index f5766f9a5..000000000
--- a/examples/exp_pool/manager.py
+++ /dev/null
@@ -1,21 +0,0 @@
-from metagpt.exp_pool.manager import ExperiencePoolManager
-from metagpt.exp_pool.schema import Experience
-from pprint import pprint
-import asyncio
-# import logging
-# logging.basicConfig(level=logging.DEBUG)
-
-async def main():
-    req = "2048 game"
-    exp = Experience(req=req, resp="python code")
-
-    manager = ExperiencePoolManager()
-
-    # pprint(manager.storage.get())
-    # manager.create_exp(exp)
-    result = await manager.query_exp(req)
-    print(result)
-
-
-if __name__ == "__main__":
-    asyncio.run(main())
\ No newline at end of file
diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/simple.py
new file mode 100644
index 000000000..bc20fbcdd
--- /dev/null
+++ b/examples/exp_pool/simple.py
@@ -0,0 +1,29 @@
+"""Simple example of experience pool."""
+
+import asyncio
+
+from metagpt.exp_pool import exp_manager
+from metagpt.exp_pool.schema import EntryType, Experience
+from metagpt.logs import logger
+
+
+async def main():
+    req = "Simple task."
+
+    # 1. Find experiences.
+    exps = await exp_manager.query_exps(req)
+    if exps:
+        logger.info(f"Experiences already exist for the request `{req}`: {exps}")
+        return
+
+    # 2. Create a new experience if none exist
+    exp_manager.create_exp(Experience(req=req, resp="Simple echo.", entry_type=EntryType.MANUAL))
+    logger.info(f"New experience created for the request `{req}`.")
+
+    # 3. Find again
+    exps = await exp_manager.query_exps(req)
+    logger.info(f"Updated experiences: {exps}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/metagpt/config2.py b/metagpt/config2.py
index 6f5a1add6..6588a6036 100644
--- a/metagpt/config2.py
+++ b/metagpt/config2.py
@@ -13,6 +13,7 @@ from pydantic import BaseModel, model_validator
 
 from metagpt.configs.browser_config import BrowserConfig
 from metagpt.configs.embedding_config import EmbeddingConfig
+from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 from metagpt.configs.llm_config import LLMConfig, LLMType
 from metagpt.configs.mermaid_config import MermaidConfig
 from metagpt.configs.redis_config import RedisConfig
@@ -22,7 +23,6 @@ from metagpt.configs.search_config import SearchConfig
 from metagpt.configs.workspace_config import WorkspaceConfig
 from metagpt.const import CONFIG_ROOT, METAGPT_ROOT
 from metagpt.utils.yaml_model import YamlModel
-from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 
 
 class CLIParams(BaseModel):
@@ -73,7 +73,7 @@ class Config(CLIParams, YamlModel):
     code_review_k_times: int = 2
 
     # Experience Pool Parameters
-    experience_pool: Optional[ExperiencePoolConfig] = None
+    exp_pool: ExperiencePoolConfig = ExperiencePoolConfig()
 
     # Will be removed in the future
     metagpt_tti_url: str = ""
diff --git a/metagpt/configs/exp_pool_config.py b/metagpt/configs/exp_pool_config.py
index f7312d2de..3f86173c1 100644
--- a/metagpt/configs/exp_pool_config.py
+++ b/metagpt/configs/exp_pool_config.py
@@ -1,6 +1,8 @@
+from pydantic import Field
+
 from metagpt.utils.yaml_model import YamlModel
 
 
 class ExperiencePoolConfig(YamlModel):
-    enable_read: bool = False
-    enable_write: bool = False
+    enable_read: bool = Field(default=True, description="Enable to read from experience pool.")
+    enable_write: bool = Field(default=True, description="Enable to write to experience pool.")
diff --git a/metagpt/exp_pool/__init__.py b/metagpt/exp_pool/__init__.py
index e69de29bb..aeeb94b38 100644
--- a/metagpt/exp_pool/__init__.py
+++ b/metagpt/exp_pool/__init__.py
@@ -0,0 +1,6 @@
+"""Experience pool init."""
+
+from metagpt.exp_pool.manager import exp_manager
+from metagpt.exp_pool.decorator import exp_cache
+
+__all__ = ["exp_manager", "exp_cache"]
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 6629e8377..1d691b8f3 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -1,4 +1,56 @@
+"""Experience Decorator."""
+
+import asyncio
+import functools
+from typing import Any, Callable, Optional, TypeVar
+
+from metagpt.exp_pool.manager import exp_manager
+from metagpt.exp_pool.schema import Experience
+from metagpt.utils.async_helper import NestAsyncio
+
+ReturnType = TypeVar("ReturnType")
 
 
-def exp_cache(func):
-    pass
+def exp_cache(_func: Optional[Callable[..., ReturnType]] = None):
+    """Decorator to check for a perfect experience and returns it if exists.
+
+    Otherwise, it executes the function, save the result as a new experience, and returns the result.
+
+    This can be applied to both synchronous and asynchronous functions.
+    """
+
+    def decorator(func: Callable[..., ReturnType]) -> Callable[..., ReturnType]:
+        @functools.wraps(func)
+        async def get_or_create(args: Any, kwargs: Any, is_async: bool) -> ReturnType:
+            """Attempts to retrieve a cached experience or creates one if not found."""
+
+            req = f"{func.__name__}_{args}_{kwargs}"
+            exps = await exp_manager.query_exps(req)
+            if perfect_exp := exp_manager.extract_one_perfect_exp(exps):
+                return perfect_exp
+
+            if is_async:
+                result = await func(*args, **kwargs)
+            else:
+                result = func(*args, **kwargs)
+
+            exp_manager.create_exp(Experience(req=req, resp=result))
+
+            return result
+
+        def sync_wrapper(*args: Any, **kwargs: Any) -> ReturnType:
+            NestAsyncio.apply_once()
+            return asyncio.get_event_loop().run_until_complete(get_or_create(args, kwargs, is_async=False))
+
+        async def async_wrapper(*args: Any, **kwargs: Any) -> ReturnType:
+            return await get_or_create(args, kwargs, is_async=True)
+
+        if asyncio.iscoroutinefunction(func):
+            return async_wrapper
+        else:
+            return sync_wrapper
+
+    if _func is None:
+        return decorator
+    else:
+        return decorator(_func)
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index c32073a9f..4bc566104 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -1,32 +1,105 @@
-from pydantic import BaseModel, ConfigDict
-from metagpt.exp_pool.schema import Experience
-import uuid
-import chromadb
-from chromadb import Collection, QueryResult
+"""Experience Manager."""
+
 from typing import Optional
+
+from pydantic import BaseModel, ConfigDict, model_validator
+
+from metagpt.config2 import Config, config
+from metagpt.exp_pool.schema import MAX_SCORE, Experience
 from metagpt.rag.engines import SimpleEngine
-from metagpt.rag.schema import ChromaRetrieverConfig
+from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
 
 
-class ExperiencePoolManager(BaseModel):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self._storage = None
+class ExperienceManager(BaseModel):
+    """ExperienceManager manages the lifecycle of experiences, including CRUD and optimization.
+
+    Attributes:
+        config (Config): Configuration for managing experiences.
+        storage (SimpleEngine): Engine to handle the storage and retrieval of experiences.
+    """
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    config: Config = config
+    storage: SimpleEngine = None
+
+    @model_validator(mode="after")
+    def initialize(self):
+        if self.storage is None:
+            self.storage = SimpleEngine.from_objs(
+                retriever_configs=[
+                    ChromaRetrieverConfig(collection_name="experience_pool", persist_path=".chroma_exp_data")
+                ],
+                ranker_configs=[LLMRankerConfig()],
+            )
+        return self
 
-    @property
-    def storage(self) -> SimpleEngine:
-        if self._storage is None:
-            self._storage = SimpleEngine.from_objs(retriever_configs=[ChromaRetrieverConfig(collection_name="experience_pool", persist_path="./chroma_data")])
-        return self._storage
-    
     def create_exp(self, exp: Experience):
+        """Adds an experience to the storage if writing is enabled.
+
+        Args:
+            exp (Experience): The experience to add.
+        """
+        if not self.config.exp_pool.enable_write:
+            return
+
         self.storage.add_objs([exp])
-    
-    async def query_exp(self, req: str) -> list[Experience]:
+
+    async def query_exps(self, req: str, tag: str = "") -> list[Experience]:
+        """Retrieves and filters experiences.
+
+        Args:
+            req (str): The query string to retrieve experiences.
+            tag (str): Optional tag to filter the experiences by.
+
+        Returns:
+            list[Experience]: A list of experiences that match the args.
+        """
+        if not self.config.exp_pool.enable_read:
+            return []
+
         nodes = await self.storage.aretrieve(req)
-        exps = [node.metadata["obj"] for node in nodes]
+        exps: list[Experience] = [node.metadata["obj"] for node in nodes]
+
+        # TODO: filter by metadata
+        if tag:
+            exps = [exp for exp in exps if exp.tag == tag]
 
         return exps
 
-    
+    def extract_one_perfect_exp(self, exps: list[Experience]) -> Optional[Experience]:
+        """Extracts the first 'perfect' experience from a list of experiences.
 
+        Args:
+            exps (list[Experience]): The experiences to evaluate.
+
+        Returns:
+            Optional[Experience]: The first perfect experience if found, otherwise None.
+        """
+        for exp in exps:
+            if self.is_perfect_exp(exp):
+                return exp
+
+        return None
+
+    @staticmethod
+    def is_perfect_exp(exp: Experience) -> bool:
+        """Determines if an experience is considered 'perfect'.
+
+        Args:
+            exp (Experience): The experience to evaluate.
+
+        Returns:
+            bool: True if the experience is manually entered, otherwise False.
+        """
+        if not exp:
+            return False
+
+        # TODO: need more metrics
+        if exp.metric and exp.metric.score == MAX_SCORE:
+            return True
+
+        return False
+
+
+exp_manager = ExperienceManager()
diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index 359268612..b51bc3c17 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -1,10 +1,46 @@
-from pydantic import BaseModel, Field
+"""Experience schema."""
+
+from enum import Enum
+from typing import Optional
+
 from llama_index.core.schema import TextNode
+from pydantic import BaseModel, Field
+
+MAX_SCORE = 10
+
+
+class ExperienceType(str, Enum):
+    """Experience Type."""
+
+    SUCCESS = "success"
+    FAILURE = "failure"
+    INSIGHT = "insight"
+
+
+class EntryType(Enum):
+    """Experience Entry Type."""
+
+    AUTOMATIC = "Automatic"
+    MANUAL = "Manual"
+
+
+class Metric(BaseModel):
+    """Experience Metric."""
+
+    time_cost: float = Field(default=0.000, description="Time cost, the unit is milliseconds.")
+    money_cost: float = Field(default=0.000, description="Money cost, the unit is US dollars.")
+    score: int = Field(default=1, description="Score, a value between 1 and 10.")
 
 
 class Experience(BaseModel):
+    """Experience."""
+
     req: str = Field(..., description="")
-    resp: str = Field(..., description="")
+    resp: str = Field(..., description="The type is string/json/code.")
+    metric: Optional[Metric] = Field(default=None, description="Metric.")
+    exp_type: ExperienceType = Field(default=ExperienceType.SUCCESS, description="The type of experience.")
+    entry_type: EntryType = Field(default=EntryType.AUTOMATIC, description="Type of entry: Manual or Automatic.")
+    tag: str = Field(default="", description="Tagging experience.")
 
     def rag_key(self):
         return self.req
diff --git a/metagpt/utils/reflection.py b/metagpt/utils/reflection.py
index 688831f06..2683e5657 100644
--- a/metagpt/utils/reflection.py
+++ b/metagpt/utils/reflection.py
@@ -23,5 +23,5 @@ def get_func_full_name(func, *args) -> str:
     if inspect.ismethod(func) or (inspect.isfunction(func) and "self" in inspect.signature(func).parameters):
         cls_name = args[0].__class__.__name__
         return f"{func.__module__}.{cls_name}.{func.__name__}"
-    
+
     return f"{func.__module__}.{func.__name__}"

From d1198dc58de9e5503e1bdbc42297e25957c83ea0 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 11:33:34 +0800
Subject: [PATCH 17/51] experiment pool init

---
 examples/exp_pool/simple.py | 13 +++----------
 1 file changed, 3 insertions(+), 10 deletions(-)

diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/simple.py
index bc20fbcdd..608578519 100644
--- a/examples/exp_pool/simple.py
+++ b/examples/exp_pool/simple.py
@@ -9,20 +9,13 @@ from metagpt.logs import logger
 
 async def main():
     req = "Simple task."
+    resp = "Simple echo."
 
-    # 1. Find experiences.
-    exps = await exp_manager.query_exps(req)
-    if exps:
-        logger.info(f"Experiences already exist for the request `{req}`: {exps}")
-        return
-
-    # 2. Create a new experience if none exist
-    exp_manager.create_exp(Experience(req=req, resp="Simple echo.", entry_type=EntryType.MANUAL))
+    exp_manager.create_exp(Experience(req=req, resp=resp, entry_type=EntryType.MANUAL))
     logger.info(f"New experience created for the request `{req}`.")
 
-    # 3. Find again
     exps = await exp_manager.query_exps(req)
-    logger.info(f"Updated experiences: {exps}")
+    logger.info(f"Got experiences: {exps}")
 
 
 if __name__ == "__main__":

From 8edfa02533a41002723d3cf5e871204e9e980f16 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 11:46:57 +0800
Subject: [PATCH 18/51] experiment pool init

---
 examples/exp_pool/simple.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/simple.py
index 608578519..f270824bf 100644
--- a/examples/exp_pool/simple.py
+++ b/examples/exp_pool/simple.py
@@ -10,8 +10,9 @@ from metagpt.logs import logger
 async def main():
     req = "Simple task."
     resp = "Simple echo."
+    exp = Experience(req=req, resp=resp, entry_type=EntryType.MANUAL)
 
-    exp_manager.create_exp(Experience(req=req, resp=resp, entry_type=EntryType.MANUAL))
+    exp_manager.create_exp(exp)
     logger.info(f"New experience created for the request `{req}`.")
 
     exps = await exp_manager.query_exps(req)

From 96cd6b5f64f0e3f4cbc4bd37fb97f08c3e4d197b Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 15:16:58 +0800
Subject: [PATCH 19/51] add trajectory schema

---
 metagpt/exp_pool/schema.py | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index b51bc3c17..e6ae4ee1d 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -32,6 +32,14 @@ class Metric(BaseModel):
     score: int = Field(default=1, description="Score, a value between 1 and 10.")
 
 
+class Trajectory(BaseModel):
+    """Experience Trajectory."""
+
+    plan: str = Field(default="", description="The plan.")
+    action: str = Field(default="", description="Action for the plan.")
+    observation: str = Field(default="", description="Output of the action.")
+
+
 class Experience(BaseModel):
     """Experience."""
 
@@ -41,6 +49,7 @@ class Experience(BaseModel):
     exp_type: ExperienceType = Field(default=ExperienceType.SUCCESS, description="The type of experience.")
     entry_type: EntryType = Field(default=EntryType.AUTOMATIC, description="Type of entry: Manual or Automatic.")
     tag: str = Field(default="", description="Tagging experience.")
+    traj: Optional[Trajectory] = Field(default=None, description="Trajectory.")
 
     def rag_key(self):
         return self.req

From d600cc47f45f07d144c65669959530bbba4069c7 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 4 Jun 2024 22:08:40 +0800
Subject: [PATCH 20/51] add exp_pool test

---
 metagpt/exp_pool/schema.py             |  5 +-
 tests/metagpt/exp_pool/test_manager.py | 77 ++++++++++++++++++++++++++
 2 files changed, 80 insertions(+), 2 deletions(-)
 create mode 100644 tests/metagpt/exp_pool/test_manager.py

diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index e6ae4ee1d..1afcc1508 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -1,7 +1,7 @@
 """Experience schema."""
 
 from enum import Enum
-from typing import Optional
+from typing import Any, Optional
 
 from llama_index.core.schema import TextNode
 from pydantic import BaseModel, Field
@@ -38,13 +38,14 @@ class Trajectory(BaseModel):
     plan: str = Field(default="", description="The plan.")
     action: str = Field(default="", description="Action for the plan.")
     observation: str = Field(default="", description="Output of the action.")
+    reward: int = Field(default=0, description="Measure the action.")
 
 
 class Experience(BaseModel):
     """Experience."""
 
     req: str = Field(..., description="")
-    resp: str = Field(..., description="The type is string/json/code.")
+    resp: Any = Field(..., description="The type is string/json/code.")
     metric: Optional[Metric] = Field(default=None, description="Metric.")
     exp_type: ExperienceType = Field(default=ExperienceType.SUCCESS, description="The type of experience.")
     entry_type: EntryType = Field(default=EntryType.AUTOMATIC, description="Type of entry: Manual or Automatic.")
diff --git a/tests/metagpt/exp_pool/test_manager.py b/tests/metagpt/exp_pool/test_manager.py
new file mode 100644
index 000000000..a0d7005f5
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_manager.py
@@ -0,0 +1,77 @@
+import pytest
+
+from metagpt.config2 import Config
+from metagpt.configs.exp_pool_config import ExperiencePoolConfig
+from metagpt.configs.llm_config import LLMConfig
+from metagpt.exp_pool.manager import ExperienceManager
+from metagpt.exp_pool.schema import MAX_SCORE, Experience, Metric
+from metagpt.rag.engines import SimpleEngine
+
+
+class TestExperienceManager:
+    @pytest.fixture
+    def mock_config(self):
+        return Config(llm=LLMConfig(), exp_pool=ExperiencePoolConfig(enable_write=True, enable_read=True))
+
+    @pytest.fixture
+    def mock_storage(self, mocker):
+        engine = mocker.MagicMock(spec=SimpleEngine)
+        engine.add_objs = mocker.MagicMock()
+        engine.aretrieve = mocker.AsyncMock(return_value=[])
+        return engine
+
+    @pytest.fixture
+    def mock_experience_manager(self, mock_config, mock_storage):
+        return ExperienceManager(config=mock_config, storage=mock_storage)
+
+    @pytest.fixture
+    def mock_experience(self):
+        return Experience(req="req", resp="resp")
+
+    def test_initialize_storage(self, mock_experience_manager, mock_storage):
+        assert mock_experience_manager.storage is mock_storage
+
+    def test_create_exp(self, mock_experience_manager, mock_experience):
+        mock_experience_manager.create_exp(mock_experience)
+        mock_experience_manager.storage.add_objs.assert_called_once_with([mock_experience])
+
+    def test_create_exp_write_disabled(self, mock_experience_manager, mock_experience, mock_config):
+        mock_config.exp_pool.enable_write = False
+        mock_experience_manager.create_exp(mock_experience)
+        mock_experience_manager.storage.add_objs.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_query_exps(self, mock_experience_manager, mocker):
+        req = "req"
+        resp = "resp"
+        tag = "test"
+        experiences = [Experience(req=req, resp=resp, tag="test"), Experience(req=req, resp=resp, tag="other")]
+        mock_experience_manager.storage.aretrieve.return_value = [
+            mocker.MagicMock(metadata={"obj": exp}) for exp in experiences
+        ]
+
+        result = await mock_experience_manager.query_exps(req, tag)
+        assert len(result) == 1
+        assert result[0].tag == "test"
+
+    @pytest.mark.asyncio
+    async def test_query_exps_no_read_permission(self, mock_experience_manager, mock_config):
+        mock_config.exp_pool.enable_read = False
+        result = await mock_experience_manager.query_exps("query")
+        assert result == []
+
+    def test_extract_one_perfect_exp(self, mock_experience_manager):
+        experiences = [
+            Experience(req="req", resp="resp", metric=Metric(score=MAX_SCORE)),
+            Experience(req="req", resp="resp"),
+        ]
+        perfect_exp: Experience = mock_experience_manager.extract_one_perfect_exp(experiences)
+        assert perfect_exp is not None
+        assert perfect_exp.metric.score == MAX_SCORE
+
+    def test_is_perfect_exp(self):
+        exp = Experience(req="req", resp="resp", metric=Metric(score=MAX_SCORE))
+        assert ExperienceManager.is_perfect_exp(exp) == True
+
+        exp = Experience(req="req", resp="resp")
+        assert ExperienceManager.is_perfect_exp(exp) == False

From 9b65dea952d788fa9632379e6806d66a1b4f3a5b Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Wed, 5 Jun 2024 10:32:48 +0800
Subject: [PATCH 21/51] update comment

---
 metagpt/exp_pool/decorator.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 1d691b8f3..e073ee494 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -22,18 +22,21 @@ def exp_cache(_func: Optional[Callable[..., ReturnType]] = None):
     def decorator(func: Callable[..., ReturnType]) -> Callable[..., ReturnType]:
         @functools.wraps(func)
         async def get_or_create(args: Any, kwargs: Any, is_async: bool) -> ReturnType:
-            """Attempts to retrieve a cached experience or creates one if not found."""
+            """Attempts to retrieve a perfect experience or creates an experience if not found."""
 
+            # 1. Get exps.
             req = f"{func.__name__}_{args}_{kwargs}"
             exps = await exp_manager.query_exps(req)
             if perfect_exp := exp_manager.extract_one_perfect_exp(exps):
                 return perfect_exp
 
+            # 2. Exec func. TODO: pass exps to func
             if is_async:
                 result = await func(*args, **kwargs)
             else:
                 result = func(*args, **kwargs)
 
+            # 3. Create an exp.
             exp_manager.create_exp(Experience(req=req, resp=result))
 
             return result

From 07bf103fb08c7c0075079380dc2692accc03e1a5 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Wed, 5 Jun 2024 22:15:09 +0800
Subject: [PATCH 22/51] add exp_pool tests

---
 examples/exp_pool/decorator.py           |   5 +-
 metagpt/exp_pool/decorator.py            | 144 ++++++++++++++++------
 metagpt/exp_pool/manager.py              |  10 +-
 metagpt/exp_pool/schema.py               |  16 ++-
 metagpt/exp_pool/scorers/__init__.py     |   6 +
 metagpt/exp_pool/scorers/base.py         |  27 +++++
 metagpt/exp_pool/scorers/simple.py       |  73 ++++++++++++
 tests/metagpt/exp_pool/test_decorator.py | 145 +++++++++++++++++++++++
 tests/metagpt/exp_pool/test_manager.py   |   8 +-
 9 files changed, 391 insertions(+), 43 deletions(-)
 create mode 100644 metagpt/exp_pool/scorers/__init__.py
 create mode 100644 metagpt/exp_pool/scorers/base.py
 create mode 100644 metagpt/exp_pool/scorers/simple.py
 create mode 100644 tests/metagpt/exp_pool/test_decorator.py

diff --git a/examples/exp_pool/decorator.py b/examples/exp_pool/decorator.py
index 2f6397f80..3f6093e01 100644
--- a/examples/exp_pool/decorator.py
+++ b/examples/exp_pool/decorator.py
@@ -7,8 +7,9 @@ from metagpt.exp_pool import exp_cache, exp_manager
 from metagpt.logs import logger
 
 
-@exp_cache
-async def produce(req):
+@exp_cache(pass_exps_to_func=True)
+async def produce(req, exps=None):
+    logger.info(f"Previous experiences: {exps}")
     return f"{req} {uuid.uuid4().hex}"
 
 
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index e073ee494..9eb4d9e61 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -4,56 +4,134 @@ import asyncio
 import functools
 from typing import Any, Callable, Optional, TypeVar
 
-from metagpt.exp_pool.manager import exp_manager
-from metagpt.exp_pool.schema import Experience
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.exp_pool.manager import ExperienceManager, exp_manager
+from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
+from metagpt.exp_pool.scorers import ExperienceScorer, SimpleScorer
 from metagpt.utils.async_helper import NestAsyncio
 
 ReturnType = TypeVar("ReturnType")
 
 
-def exp_cache(_func: Optional[Callable[..., ReturnType]] = None):
-    """Decorator to check for a perfect experience and returns it if exists.
-
-    Otherwise, it executes the function, save the result as a new experience, and returns the result.
+def exp_cache(
+    _func: Optional[Callable[..., ReturnType]] = None,
+    query_type: QueryType = QueryType.SEMANTIC,
+    scorer: Optional[ExperienceScorer] = None,
+    manager: Optional[ExperienceManager] = None,
+    pass_exps_to_func: bool = False,
+):
+    """Decorator to get a perfect experience, otherwise, it executes the function, and create a new experience.
 
     This can be applied to both synchronous and asynchronous functions.
+
+    Args:
+        _func: Just to make the decorator more flexible, for example, it can be used directly with @exp_cache by default, without the need for @exp_cache().
+        query_type: The type of query to be used when fetching experiences.
+        scorer: Evaluate experience. Default SimpleScorer.
+        manager: How to fetch, evaluate and save experience, etc. Default exp_manager.
+        pass_exps_to_func: To control whether imperfect experiences are passed to the function, if True, the func must have a parameter named 'exps'.
     """
 
     def decorator(func: Callable[..., ReturnType]) -> Callable[..., ReturnType]:
         @functools.wraps(func)
-        async def get_or_create(args: Any, kwargs: Any, is_async: bool) -> ReturnType:
-            """Attempts to retrieve a perfect experience or creates an experience if not found."""
+        async def get_or_create(args: Any, kwargs: Any) -> ReturnType:
+            handler = ExpCacheHandler(
+                func=func,
+                args=args,
+                kwargs=kwargs,
+                exp_manager=manager or exp_manager,
+                exp_scorer=scorer or SimpleScorer(),
+                pass_exps=pass_exps_to_func,
+            )
 
-            # 1. Get exps.
-            req = f"{func.__name__}_{args}_{kwargs}"
-            exps = await exp_manager.query_exps(req)
-            if perfect_exp := exp_manager.extract_one_perfect_exp(exps):
-                return perfect_exp
+            await handler.fetch_experiences(query_type)
+            if exp := handler.get_one_perfect_experience():
+                return exp
 
-            # 2. Exec func. TODO: pass exps to func
-            if is_async:
-                result = await func(*args, **kwargs)
-            else:
-                result = func(*args, **kwargs)
+            await handler.execute_function()
+            await handler.evaluate_experience()
+            handler.save_experience()
 
-            # 3. Create an exp.
-            exp_manager.create_exp(Experience(req=req, resp=result))
+            return handler._result
 
-            return result
+        return ExpCacheHandler.choose_wrapper(func, get_or_create)
 
-        def sync_wrapper(*args: Any, **kwargs: Any) -> ReturnType:
+    return decorator(_func) if _func else decorator
+
+
+class ExpCacheHandler(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    func: Callable
+    args: Any
+    kwargs: Any
+    exp_manager: ExperienceManager
+    exp_scorer: ExperienceScorer
+    pass_exps: bool
+
+    _exps: list[Experience] = None
+    _result: Any = None
+    _score: Score = None
+
+    async def fetch_experiences(self, query_type: QueryType):
+        """Fetch a potentially perfect existing experience."""
+
+        req = self.generate_req_identifier()
+        self._exps = await self.exp_manager.query_exps(req, query_type=query_type)
+
+    def get_one_perfect_experience(self) -> Optional[Experience]:
+        return self.exp_manager.extract_one_perfect_exp(self._exps)
+
+    async def execute_function(self):
+        """Execute the function, and save the result."""
+        self._result = await self._execute_function()
+
+    async def evaluate_experience(self):
+        """Evaluate the experience, and save the score."""
+
+        self._score = await self.exp_scorer.evaluate(self.func, self._result, self.args, self.kwargs)
+
+    def save_experience(self):
+        """Save the new experience."""
+
+        req = self.generate_req_identifier()
+        exp = Experience(req=req, resp=self._result, metric=Metric(score=self._score))
+
+        self.exp_manager.create_exp(exp)
+
+    def generate_req_identifier(self):
+        """Generate a unique request identifier based on the function and its arguments."""
+
+        return f"{self.func.__name__}_{self.args}_{self.kwargs}"
+
+    @staticmethod
+    def choose_wrapper(func, wrapped_func):
+        """Choose how to run wrapped_func based on whether the function is asynchronous."""
+
+        async def async_wrapper(*args, **kwargs):
+            return await wrapped_func(args, kwargs)
+
+        def sync_wrapper(*args, **kwargs):
             NestAsyncio.apply_once()
-            return asyncio.get_event_loop().run_until_complete(get_or_create(args, kwargs, is_async=False))
+            return asyncio.get_event_loop().run_until_complete(wrapped_func(args, kwargs))
 
-        async def async_wrapper(*args: Any, **kwargs: Any) -> ReturnType:
-            return await get_or_create(args, kwargs, is_async=True)
+        return async_wrapper if asyncio.iscoroutinefunction(func) else sync_wrapper
 
-        if asyncio.iscoroutinefunction(func):
-            return async_wrapper
-        else:
-            return sync_wrapper
+    async def _execute_function(self):
+        if self.pass_exps:
+            return await self._execute_function_with_exps()
 
-    if _func is None:
-        return decorator
-    else:
-        return decorator(_func)
+        return await self._execute_function_without_exps()
+
+    async def _execute_function_without_exps(self):
+        if asyncio.iscoroutinefunction(self.func):
+            return await self.func(*self.args, **self.kwargs)
+
+        return self.func(*self.args, **self.kwargs)
+
+    async def _execute_function_with_exps(self):
+        if asyncio.iscoroutinefunction(self.func):
+            return await self.func(*self.args, **self.kwargs, exps=self._exps)
+
+        return self.func(*self.args, **self.kwargs, exps=self._exps)
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 4bc566104..58499104d 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -5,7 +5,7 @@ from typing import Optional
 from pydantic import BaseModel, ConfigDict, model_validator
 
 from metagpt.config2 import Config, config
-from metagpt.exp_pool.schema import MAX_SCORE, Experience
+from metagpt.exp_pool.schema import MAX_SCORE, Experience, QueryType
 from metagpt.rag.engines import SimpleEngine
 from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
 
@@ -45,12 +45,13 @@ class ExperienceManager(BaseModel):
 
         self.storage.add_objs([exp])
 
-    async def query_exps(self, req: str, tag: str = "") -> list[Experience]:
+    async def query_exps(self, req: str, tag: str = "", query_type: QueryType = QueryType.SEMANTIC) -> list[Experience]:
         """Retrieves and filters experiences.
 
         Args:
             req (str): The query string to retrieve experiences.
             tag (str): Optional tag to filter the experiences by.
+            query_type (QueryType): Default semantic to vector matching. exact to same matching.
 
         Returns:
             list[Experience]: A list of experiences that match the args.
@@ -65,6 +66,9 @@ class ExperienceManager(BaseModel):
         if tag:
             exps = [exp for exp in exps if exp.tag == tag]
 
+        if query_type == QueryType.EXACT:
+            exps = [exp for exp in exps if exp.req == req]
+
         return exps
 
     def extract_one_perfect_exp(self, exps: list[Experience]) -> Optional[Experience]:
@@ -96,7 +100,7 @@ class ExperienceManager(BaseModel):
             return False
 
         # TODO: need more metrics
-        if exp.metric and exp.metric.score == MAX_SCORE:
+        if exp.metric and exp.metric.score.val == MAX_SCORE:
             return True
 
         return False
diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index 1afcc1508..9fc665cca 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -9,6 +9,13 @@ from pydantic import BaseModel, Field
 MAX_SCORE = 10
 
 
+class QueryType(str, Enum):
+    """Type of query experiences."""
+
+    EXACT = "exact"
+    SEMANTIC = "semantic"
+
+
 class ExperienceType(str, Enum):
     """Experience Type."""
 
@@ -24,12 +31,19 @@ class EntryType(Enum):
     MANUAL = "Manual"
 
 
+class Score(BaseModel):
+    """Score in Metric."""
+
+    val: int = Field(default=1, description="Value of the score, Between 1 and 10, higher is better.")
+    reason: str = Field(default="", description="Reason for the value.")
+
+
 class Metric(BaseModel):
     """Experience Metric."""
 
     time_cost: float = Field(default=0.000, description="Time cost, the unit is milliseconds.")
     money_cost: float = Field(default=0.000, description="Money cost, the unit is US dollars.")
-    score: int = Field(default=1, description="Score, a value between 1 and 10.")
+    score: Score = Field(default=None, description="Score, with value and reason.")
 
 
 class Trajectory(BaseModel):
diff --git a/metagpt/exp_pool/scorers/__init__.py b/metagpt/exp_pool/scorers/__init__.py
new file mode 100644
index 000000000..85bea88ff
--- /dev/null
+++ b/metagpt/exp_pool/scorers/__init__.py
@@ -0,0 +1,6 @@
+"""Experience scorers init."""
+
+from metagpt.exp_pool.scorers.base import ExperienceScorer
+from metagpt.exp_pool.scorers.simple import SimpleScorer
+
+__all__ = ["ExperienceScorer", "SimpleScorer"]
diff --git a/metagpt/exp_pool/scorers/base.py b/metagpt/exp_pool/scorers/base.py
new file mode 100644
index 000000000..a9d30cffe
--- /dev/null
+++ b/metagpt/exp_pool/scorers/base.py
@@ -0,0 +1,27 @@
+"""Experience Scorers."""
+
+from abc import abstractmethod
+from typing import Any, Callable
+
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.exp_pool.schema import Score
+
+
+class ExperienceScorer(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    @abstractmethod
+    async def evaluate(self, func: Callable, result: Any, args: tuple = None, kwargs: dict = None) -> Score:
+        """Evaluate the quality of the result produced by the function and parameters.
+
+        Args:
+            func (Callable): The function whose result is to be evaluated.
+            result (Any): The result produced by the function.
+            args (Tuple[Any, ...]): The tuple of arguments that were passed to the function.
+            kwargs (Dict[str, Any]): The dictionary of keyword arguments that were passed to the function.
+
+        Example:
+            result = await sample(5, name="foo")
+            score = await scorer.evaluate(sample, result, args=(5), kwargs={"name": "foo"})
+        """
diff --git a/metagpt/exp_pool/scorers/simple.py b/metagpt/exp_pool/scorers/simple.py
new file mode 100644
index 000000000..d0301cbc2
--- /dev/null
+++ b/metagpt/exp_pool/scorers/simple.py
@@ -0,0 +1,73 @@
+"""Evalate by llm."""
+import inspect
+import json
+from typing import Any, Callable
+
+from pydantic import Field
+
+from metagpt.exp_pool.schema import Score
+from metagpt.exp_pool.scorers.base import ExperienceScorer
+from metagpt.llm import LLM
+from metagpt.provider.base_llm import BaseLLM
+from metagpt.utils.common import parse_json_code_block
+
+SIMPLE_SCORER_TEMPLATE = """
+Role: You're an expert score evaluator. You specialize in assessing the output of the given function, based on its intended requirement and produced result.
+
+## Context
+### Function Name
+{func_name}
+
+### Function Document
+{func_doc}
+
+### Function Signature
+{func_signature}
+
+### Function Parameters
+args: {func_args}
+kwargs: {func_kwargs}
+
+### Produced Result By Function and Parameters
+{func_result}
+
+## Format Example
+```json
+{{
+    "val": "the value of the score, int from 1 to 10, higher is better.",
+    "reason": "an explanation supporting the score."
+}}
+```
+
+## Instructions
+- Understand the function and requirements given by the user.
+- Analyze the results produced by the function.
+- Grade the results based on level of alignment with the requirements.
+- Provide a score on a scale defined by user or a default scale (1 to 10).
+
+## Constraint
+Format: Just print the result in json format like **Format Example**.
+
+## Action
+Follow instructions, generate output and make sure it follows the **Constraint**.
+"""
+
+
+class SimpleScorer(ExperienceScorer):
+    llm: BaseLLM = Field(default_factory=LLM)
+
+    async def evaluate(self, func: Callable, result: Any, args: tuple = None, kwargs: dict = None) -> Score:
+        """Evaluate the quality of content."""
+
+        prompt = SIMPLE_SCORER_TEMPLATE.format(
+            func_name=func.__name__,
+            func_doc=func.__doc__,
+            func_signature=inspect.signature(func),
+            func_args=args,
+            func_kwargs=kwargs,
+            func_result=result,
+        )
+        resp = await self.llm.aask(prompt)
+        resp_json = json.loads(parse_json_code_block(resp)[0])
+
+        return Score(**resp_json)
diff --git a/tests/metagpt/exp_pool/test_decorator.py b/tests/metagpt/exp_pool/test_decorator.py
new file mode 100644
index 000000000..508229d18
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_decorator.py
@@ -0,0 +1,145 @@
+import asyncio
+
+import pytest
+
+from metagpt.exp_pool.decorator import ExpCacheHandler
+from metagpt.exp_pool.manager import ExperienceManager
+from metagpt.exp_pool.schema import Experience, QueryType, Score
+from metagpt.exp_pool.scorers import SimpleScorer
+from metagpt.rag.engines import SimpleEngine
+
+
+class TestExpCache:
+    @pytest.fixture
+    def mock_func(self, mocker):
+        return mocker.AsyncMock()
+
+    @pytest.fixture
+    def mock_exp_manager(self, mocker):
+        manager = mocker.MagicMock(spec=ExperienceManager)
+        manager.storage = mocker.MagicMock(spec=SimpleEngine)
+        manager.query_exps = mocker.AsyncMock()
+        manager.create_exp = mocker.MagicMock()
+        manager.extract_one_perfect_exp = mocker.MagicMock()
+        return manager
+
+    @pytest.fixture
+    def mock_scorer(self, mocker):
+        scorer = mocker.MagicMock(spec=SimpleScorer)
+        scorer.evaluate = mocker.AsyncMock()
+        return scorer
+
+    @pytest.fixture
+    def exp_cache_handler(self, mock_func, mock_exp_manager, mock_scorer):
+        return ExpCacheHandler(
+            func=mock_func, args=(), kwargs={}, exp_manager=mock_exp_manager, exp_scorer=mock_scorer, pass_exps=False
+        )
+
+    @pytest.mark.asyncio
+    async def test_fetch_experiences(self, exp_cache_handler, mock_exp_manager):
+        await exp_cache_handler.fetch_experiences(QueryType.SEMANTIC)
+        mock_exp_manager.query_exps.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_perfect_experience_found(self, exp_cache_handler, mock_exp_manager, mock_func):
+        # Setup: Assume perfect experience is found
+        perfect_exp = Experience(req="req", resp="resp")
+        mock_exp_manager.extract_one_perfect_exp.return_value = perfect_exp
+
+        # Execute
+        exp_cache_handler._exps = [perfect_exp]  # Simulate fetched experiences
+        result = exp_cache_handler.get_one_perfect_experience()
+
+        # Assert
+        assert result.resp == "resp"
+        mock_func.assert_not_called()  # Function should not be called
+
+    @pytest.mark.asyncio
+    async def test_execute_function_when_no_perfect_exp(self, exp_cache_handler, mock_exp_manager, mock_func):
+        # Setup: No perfect experience
+        mock_exp_manager.extract_one_perfect_exp.return_value = None
+        mock_func.return_value = "Computed result"
+
+        # Execute
+        await exp_cache_handler.execute_function()
+
+        # Assert
+        assert exp_cache_handler._result == "Computed result"
+        mock_func.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_evaluate_and_save_experience(self, exp_cache_handler, mock_scorer, mock_exp_manager):
+        # Setup
+        mock_scorer.evaluate.return_value = Score(value=100)
+        exp_cache_handler._result = "Computed result"
+
+        # Execute
+        await exp_cache_handler.evaluate_experience()
+        exp_cache_handler.save_experience()
+
+        # Assert
+        mock_scorer.evaluate.assert_called_once()
+        mock_exp_manager.create_exp.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_async_function_execution_with_exps(self, exp_cache_handler, mock_exp_manager, mock_func):
+        # Setup
+        exp_cache_handler.pass_exps = True
+        mock_func.return_value = "Async result with exps"
+        mock_exp_manager.extract_one_perfect_exp.return_value = None
+        exp_cache_handler._exps = [Experience(req="req", resp="resp")]
+
+        # Execute
+        await exp_cache_handler.execute_function()
+
+        # Assert
+        mock_func.assert_called_once_with(exps=exp_cache_handler._exps)
+        assert exp_cache_handler._result == "Async result with exps"
+
+    def test_sync_function_execution_with_exps(self, mocker, exp_cache_handler, mock_exp_manager, mock_func):
+        # Setup
+        exp_cache_handler.func = mocker.Mock(return_value="Sync result with exps")
+        exp_cache_handler.pass_exps = True
+        mock_exp_manager.extract_one_perfect_exp.return_value = None
+        exp_cache_handler._exps = [Experience(req="req", resp="resp")]
+
+        # Execute
+        asyncio.get_event_loop().run_until_complete(exp_cache_handler.execute_function())
+
+        # Assert
+        exp_cache_handler.func.assert_called_once_with(exps=exp_cache_handler._exps)
+        assert exp_cache_handler._result == "Sync result with exps"
+
+    def test_wrapper_selection_async(self, mocker, exp_cache_handler, mock_func):
+        # Setup
+        mock_func = mocker.AsyncMock()
+
+        # Execute
+        wrapper = ExpCacheHandler.choose_wrapper(mock_func, exp_cache_handler.execute_function)
+
+        # Assert
+        assert asyncio.iscoroutinefunction(wrapper), "Wrapper should be asynchronous"
+
+    def test_wrapper_selection_sync(self, exp_cache_handler, mocker):
+        # Setup
+        sync_func = mocker.Mock()
+
+        # Execute
+        wrapper = ExpCacheHandler.choose_wrapper(sync_func, exp_cache_handler.execute_function)
+
+        # Assert
+        assert not asyncio.iscoroutinefunction(wrapper), "Wrapper should be synchronous"
+
+    @pytest.mark.asyncio
+    async def test_generate_req_identifier(self, exp_cache_handler):
+        # Setup
+        exp_cache_handler.func = lambda x: x
+        exp_cache_handler.args = (42,)
+        exp_cache_handler.kwargs = {"y": 3.14}
+
+        # Execute
+        req_id = exp_cache_handler.generate_req_identifier()
+
+        # Assert
+        expected_id = "<lambda>_(42,)_{'y': 3.14}"
+        assert req_id == expected_id, "Request identifier should match the expected format"
diff --git a/tests/metagpt/exp_pool/test_manager.py b/tests/metagpt/exp_pool/test_manager.py
index a0d7005f5..3e8f47417 100644
--- a/tests/metagpt/exp_pool/test_manager.py
+++ b/tests/metagpt/exp_pool/test_manager.py
@@ -4,7 +4,7 @@ from metagpt.config2 import Config
 from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 from metagpt.configs.llm_config import LLMConfig
 from metagpt.exp_pool.manager import ExperienceManager
-from metagpt.exp_pool.schema import MAX_SCORE, Experience, Metric
+from metagpt.exp_pool.schema import MAX_SCORE, Experience, Metric, Score
 from metagpt.rag.engines import SimpleEngine
 
 
@@ -62,15 +62,15 @@ class TestExperienceManager:
 
     def test_extract_one_perfect_exp(self, mock_experience_manager):
         experiences = [
-            Experience(req="req", resp="resp", metric=Metric(score=MAX_SCORE)),
+            Experience(req="req", resp="resp", metric=Metric(score=Score(val=MAX_SCORE))),
             Experience(req="req", resp="resp"),
         ]
         perfect_exp: Experience = mock_experience_manager.extract_one_perfect_exp(experiences)
         assert perfect_exp is not None
-        assert perfect_exp.metric.score == MAX_SCORE
+        assert perfect_exp.metric.score.val == MAX_SCORE
 
     def test_is_perfect_exp(self):
-        exp = Experience(req="req", resp="resp", metric=Metric(score=MAX_SCORE))
+        exp = Experience(req="req", resp="resp", metric=Metric(score=Score(val=MAX_SCORE)))
         assert ExperienceManager.is_perfect_exp(exp) == True
 
         exp = Experience(req="req", resp="resp")

From 0510da5295b3e1183c5c7672b15095cb5c213325 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Wed, 5 Jun 2024 23:26:09 +0800
Subject: [PATCH 23/51] add handle_exception to ensure robustness

---
 metagpt/exp_pool/decorator.py | 14 ++++++++++++--
 metagpt/exp_pool/manager.py   |  3 +++
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 9eb4d9e61..9cf924779 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -10,6 +10,7 @@ from metagpt.exp_pool.manager import ExperienceManager, exp_manager
 from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
 from metagpt.exp_pool.scorers import ExperienceScorer, SimpleScorer
 from metagpt.utils.async_helper import NestAsyncio
+from metagpt.utils.exceptions import handle_exception
 
 ReturnType = TypeVar("ReturnType")
 
@@ -50,8 +51,7 @@ def exp_cache(
                 return exp
 
             await handler.execute_function()
-            await handler.evaluate_experience()
-            handler.save_experience()
+            await handler.process_experience()
 
             return handler._result
 
@@ -87,6 +87,16 @@ class ExpCacheHandler(BaseModel):
         """Execute the function, and save the result."""
         self._result = await self._execute_function()
 
+    @handle_exception
+    async def process_experience(self):
+        """Process experience.
+
+        Evaluates and saves experience.
+        Use `handle_exception` to ensure robustness, do not stop subsequent operations.
+        """
+        await self.evaluate_experience()
+        self.save_experience()
+
     async def evaluate_experience(self):
         """Evaluate the experience, and save the score."""
 
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 58499104d..546086b37 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -8,6 +8,7 @@ from metagpt.config2 import Config, config
 from metagpt.exp_pool.schema import MAX_SCORE, Experience, QueryType
 from metagpt.rag.engines import SimpleEngine
 from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
+from metagpt.utils.exceptions import handle_exception
 
 
 class ExperienceManager(BaseModel):
@@ -34,6 +35,7 @@ class ExperienceManager(BaseModel):
             )
         return self
 
+    @handle_exception
     def create_exp(self, exp: Experience):
         """Adds an experience to the storage if writing is enabled.
 
@@ -45,6 +47,7 @@ class ExperienceManager(BaseModel):
 
         self.storage.add_objs([exp])
 
+    @handle_exception(default_return=[])
     async def query_exps(self, req: str, tag: str = "", query_type: QueryType = QueryType.SEMANTIC) -> list[Experience]:
         """Retrieves and filters experiences.
 

From 6f84110e0f0aa7635e400bdf2e31346e1411b42e Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Thu, 6 Jun 2024 20:18:40 +0800
Subject: [PATCH 24/51] update exp_pool example

---
 examples/exp_pool/simple.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/simple.py
index f270824bf..3216e78b8 100644
--- a/examples/exp_pool/simple.py
+++ b/examples/exp_pool/simple.py
@@ -9,8 +9,7 @@ from metagpt.logs import logger
 
 async def main():
     req = "Simple task."
-    resp = "Simple echo."
-    exp = Experience(req=req, resp=resp, entry_type=EntryType.MANUAL)
+    exp = Experience(req=req, resp="echo", entry_type=EntryType.MANUAL)
 
     exp_manager.create_exp(exp)
     logger.info(f"New experience created for the request `{req}`.")

From faeee22dcb445ba485149d055756324a0950e5fc Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Fri, 7 Jun 2024 10:30:37 +0800
Subject: [PATCH 25/51] update comment

---
 metagpt/exp_pool/scorers/simple.py | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/metagpt/exp_pool/scorers/simple.py b/metagpt/exp_pool/scorers/simple.py
index d0301cbc2..5779f7fb1 100644
--- a/metagpt/exp_pool/scorers/simple.py
+++ b/metagpt/exp_pool/scorers/simple.py
@@ -1,4 +1,5 @@
-"""Evalate by llm."""
+"""Simple Scorer."""
+
 import inspect
 import json
 from typing import Any, Callable
@@ -57,8 +58,17 @@ class SimpleScorer(ExperienceScorer):
     llm: BaseLLM = Field(default_factory=LLM)
 
     async def evaluate(self, func: Callable, result: Any, args: tuple = None, kwargs: dict = None) -> Score:
-        """Evaluate the quality of content."""
+        """Evaluates the quality of content by LLM.
 
+        Args:
+            func: The function to evaluate.
+            result: The result produced by the function.
+            args: The positional arguments used when calling the function, if any.
+            kwargs: The keyword arguments used when calling the function, if any.
+
+        Returns:
+            A Score object containing the evaluation results.
+        """
         prompt = SIMPLE_SCORER_TEMPLATE.format(
             func_name=func.__name__,
             func_doc=func.__doc__,

From f7514632d9304798cda6d469973d53e2d00489cd Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Fri, 7 Jun 2024 14:35:47 +0800
Subject: [PATCH 26/51] update comment

---
 metagpt/exp_pool/manager.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 546086b37..35ee5fdac 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -14,7 +14,7 @@ from metagpt.utils.exceptions import handle_exception
 class ExperienceManager(BaseModel):
     """ExperienceManager manages the lifecycle of experiences, including CRUD and optimization.
 
-    Attributes:
+    Args:
         config (Config): Configuration for managing experiences.
         storage (SimpleEngine): Engine to handle the storage and retrieval of experiences.
     """

From 790ff5598192005e57c1c0da0dcc6ca887381f10 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Fri, 7 Jun 2024 18:15:23 +0800
Subject: [PATCH 27/51] add scorer example

---
 examples/exp_pool/{simple.py => manager.py} |  0
 examples/exp_pool/scorer.py                 | 25 +++++++++++++++++++++
 metagpt/exp_pool/scorers/simple.py          |  4 ++--
 3 files changed, 27 insertions(+), 2 deletions(-)
 rename examples/exp_pool/{simple.py => manager.py} (100%)
 create mode 100644 examples/exp_pool/scorer.py

diff --git a/examples/exp_pool/simple.py b/examples/exp_pool/manager.py
similarity index 100%
rename from examples/exp_pool/simple.py
rename to examples/exp_pool/manager.py
diff --git a/examples/exp_pool/scorer.py b/examples/exp_pool/scorer.py
new file mode 100644
index 000000000..1efe07bdf
--- /dev/null
+++ b/examples/exp_pool/scorer.py
@@ -0,0 +1,25 @@
+import asyncio
+
+from metagpt.exp_pool.scorers import SimpleScorer
+from metagpt.logs import logger
+
+
+def echo(req: str):
+    """Echo from req."""
+
+    return req
+
+
+async def simple():
+    scorer = SimpleScorer()
+
+    score = await scorer.evaluate(echo, "data", ("data",))
+    logger.info(f"The score is: {score}")
+
+
+async def main():
+    await simple()
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/metagpt/exp_pool/scorers/simple.py b/metagpt/exp_pool/scorers/simple.py
index 5779f7fb1..84995b60f 100644
--- a/metagpt/exp_pool/scorers/simple.py
+++ b/metagpt/exp_pool/scorers/simple.py
@@ -10,7 +10,7 @@ from metagpt.exp_pool.schema import Score
 from metagpt.exp_pool.scorers.base import ExperienceScorer
 from metagpt.llm import LLM
 from metagpt.provider.base_llm import BaseLLM
-from metagpt.utils.common import parse_json_code_block
+from metagpt.utils.common import CodeParser
 
 SIMPLE_SCORER_TEMPLATE = """
 Role: You're an expert score evaluator. You specialize in assessing the output of the given function, based on its intended requirement and produced result.
@@ -78,6 +78,6 @@ class SimpleScorer(ExperienceScorer):
             func_result=result,
         )
         resp = await self.llm.aask(prompt)
-        resp_json = json.loads(parse_json_code_block(resp)[0])
+        resp_json = json.loads(CodeParser.parse_code(resp, lang="json"))
 
         return Score(**resp_json)

From 4650b7bdf1a8eff0140ce0e6cd16245548fbeb43 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 11 Jun 2024 15:40:01 +0800
Subject: [PATCH 28/51] change req in exp

---
 metagpt/exp_pool/decorator.py          | 29 +++++++++++++++++++-------
 metagpt/exp_pool/manager.py            |  3 +--
 metagpt/utils/reflection.py            | 25 +++++++++++++++++-----
 tests/metagpt/utils/test_reflection.py | 29 ++++++++++++++++++++++++++
 4 files changed, 71 insertions(+), 15 deletions(-)
 create mode 100644 tests/metagpt/utils/test_reflection.py

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 9cf924779..e559797a3 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -11,6 +11,7 @@ from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
 from metagpt.exp_pool.scorers import ExperienceScorer, SimpleScorer
 from metagpt.utils.async_helper import NestAsyncio
 from metagpt.utils.exceptions import handle_exception
+from metagpt.utils.reflection import get_class_name
 
 ReturnType = TypeVar("ReturnType")
 
@@ -43,7 +44,7 @@ def exp_cache(
                 kwargs=kwargs,
                 exp_manager=manager or exp_manager,
                 exp_scorer=scorer or SimpleScorer(),
-                pass_exps=pass_exps_to_func,
+                pass_exps_to_func=pass_exps_to_func,
             )
 
             await handler.fetch_experiences(query_type)
@@ -68,16 +69,17 @@ class ExpCacheHandler(BaseModel):
     kwargs: Any
     exp_manager: ExperienceManager
     exp_scorer: ExperienceScorer
-    pass_exps: bool
+    pass_exps_to_func: bool = False
 
     _exps: list[Experience] = None
     _result: Any = None
     _score: Score = None
+    _req: str = None
 
     async def fetch_experiences(self, query_type: QueryType):
         """Fetch a potentially perfect existing experience."""
 
-        req = self.generate_req_identifier()
+        req = self._get_req_identifier()
         self._exps = await self.exp_manager.query_exps(req, query_type=query_type)
 
     def get_one_perfect_experience(self) -> Optional[Experience]:
@@ -105,15 +107,26 @@ class ExpCacheHandler(BaseModel):
     def save_experience(self):
         """Save the new experience."""
 
-        req = self.generate_req_identifier()
+        req = self._get_req_identifier()
         exp = Experience(req=req, resp=self._result, metric=Metric(score=self._score))
 
         self.exp_manager.create_exp(exp)
 
-    def generate_req_identifier(self):
-        """Generate a unique request identifier based on the function and its arguments."""
+    def _get_req_identifier(self):
+        """Generate a unique request identifier based on the function and its arguments.
 
-        return f"{self.func.__name__}_{self.args}_{self.kwargs}"
+        Result Example:
+        - "write_prd-('2048',)-{}"
+        - "WritePRD.run-('2048',)-{}"
+        """
+        if not self._req:
+            cls_name = get_class_name(self.func, *self.args)
+            func_name = f"{cls_name}.{self.func.__name__}" if cls_name else self.func.__name__
+            args = self.args[1:] if cls_name and len(self.args) >= 1 else self.args
+
+            self._req = f"{func_name}-{args}-{self.kwargs}"
+
+        return self._req
 
     @staticmethod
     def choose_wrapper(func, wrapped_func):
@@ -129,7 +142,7 @@ class ExpCacheHandler(BaseModel):
         return async_wrapper if asyncio.iscoroutinefunction(func) else sync_wrapper
 
     async def _execute_function(self):
-        if self.pass_exps:
+        if self.pass_exps_to_func:
             return await self._execute_function_with_exps()
 
         return await self._execute_function_without_exps()
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 35ee5fdac..7382fe8f1 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -7,7 +7,7 @@ from pydantic import BaseModel, ConfigDict, model_validator
 from metagpt.config2 import Config, config
 from metagpt.exp_pool.schema import MAX_SCORE, Experience, QueryType
 from metagpt.rag.engines import SimpleEngine
-from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
+from metagpt.rag.schema import ChromaRetrieverConfig
 from metagpt.utils.exceptions import handle_exception
 
 
@@ -31,7 +31,6 @@ class ExperienceManager(BaseModel):
                 retriever_configs=[
                     ChromaRetrieverConfig(collection_name="experience_pool", persist_path=".chroma_exp_data")
                 ],
-                ranker_configs=[LLMRankerConfig()],
             )
         return self
 
diff --git a/metagpt/utils/reflection.py b/metagpt/utils/reflection.py
index 2683e5657..9b10a4b3e 100644
--- a/metagpt/utils/reflection.py
+++ b/metagpt/utils/reflection.py
@@ -19,9 +19,24 @@ def check_methods(C, *methods):
     return True
 
 
-def get_func_full_name(func, *args) -> str:
-    if inspect.ismethod(func) or (inspect.isfunction(func) and "self" in inspect.signature(func).parameters):
-        cls_name = args[0].__class__.__name__
-        return f"{func.__module__}.{cls_name}.{func.__name__}"
+def get_class_name(func, *args) -> str:
+    """Returns the class name of the object that a method belongs to.
 
-    return f"{func.__module__}.{func.__name__}"
+    - If `func` is a bound method, extracts the class name directly from the method.
+    - If `func` is an unbound method and `args` are provided, assumes the first argument is `self` and extracts the class name.
+    - Returns an empty string if neither condition is met.
+    """
+    if inspect.ismethod(func):
+        return func.__self__.__class__.__name__
+
+    if inspect.isfunction(func) and "self" in inspect.signature(func).parameters and args:
+        return args[0].__class__.__name__
+
+    return ""
+
+
+def get_func_or_method_name(func, *args) -> str:
+    """Function name, or method name with class name."""
+    cls_name = get_class_name(func, *args)
+
+    return f"{cls_name}.{func.__name__}" if cls_name else f"{func.__name__}"
diff --git a/tests/metagpt/utils/test_reflection.py b/tests/metagpt/utils/test_reflection.py
new file mode 100644
index 000000000..e78e1b400
--- /dev/null
+++ b/tests/metagpt/utils/test_reflection.py
@@ -0,0 +1,29 @@
+from metagpt.utils.reflection import get_func_or_method_name
+
+
+def simple_function():
+    pass
+
+
+class SampleClass:
+    def method(self):
+        pass
+
+
+class TestFunctionOrMethodName:
+    def test_simple_function(self):
+        assert get_func_or_method_name(simple_function) == "simple_function"
+
+    def test_class_method_without_args(self):
+        sample_instance = SampleClass()
+        assert get_func_or_method_name(sample_instance.method) == "SampleClass.method"
+
+    def test_class_method_with_args(self):
+        sample_instance = SampleClass()
+        assert get_func_or_method_name(SampleClass.method, sample_instance) == "SampleClass.method"
+
+    def test_function_with_no_args(self):
+        assert get_func_or_method_name(simple_function) == "simple_function"
+
+    def test_method_without_instance(self):
+        assert get_func_or_method_name(SampleClass.method) == "method"

From 6052d8b9ac8095514e07dd58b4c64f46f238f693 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 11 Jun 2024 21:40:51 +0800
Subject: [PATCH 29/51] update exp_pool decorator

---
 metagpt/exp_pool/decorator.py            |  61 ++++++++----
 metagpt/exp_pool/manager.py              |   3 +-
 metagpt/utils/reflection.py              |  25 +++--
 tests/metagpt/exp_pool/test_decorator.py | 112 +++++++++++++++++++----
 tests/metagpt/utils/test_reflection.py   |  46 ++++++----
 5 files changed, 173 insertions(+), 74 deletions(-)

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index e559797a3..446220a47 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -2,9 +2,11 @@
 
 import asyncio
 import functools
+import inspect
+import json
 from typing import Any, Callable, Optional, TypeVar
 
-from pydantic import BaseModel, ConfigDict
+from pydantic import BaseModel, ConfigDict, model_validator
 
 from metagpt.exp_pool.manager import ExperienceManager, exp_manager
 from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
@@ -42,8 +44,8 @@ def exp_cache(
                 func=func,
                 args=args,
                 kwargs=kwargs,
-                exp_manager=manager or exp_manager,
-                exp_scorer=scorer or SimpleScorer(),
+                exp_manager=manager,
+                exp_scorer=scorer,
                 pass_exps_to_func=pass_exps_to_func,
             )
 
@@ -67,8 +69,8 @@ class ExpCacheHandler(BaseModel):
     func: Callable
     args: Any
     kwargs: Any
-    exp_manager: ExperienceManager
-    exp_scorer: ExperienceScorer
+    exp_manager: Optional[ExperienceManager] = None
+    exp_scorer: Optional[ExperienceScorer] = None
     pass_exps_to_func: bool = False
 
     _exps: list[Experience] = None
@@ -76,11 +78,22 @@ class ExpCacheHandler(BaseModel):
     _score: Score = None
     _req: str = None
 
+    @model_validator(mode="after")
+    def initialize(self):
+        if self.exp_manager is None:
+            self.exp_manager = exp_manager
+
+        if self.exp_scorer is None:
+            self.exp_scorer = SimpleScorer()
+
+        self._req = self.generate_req_identifier(self.func, *self.args, **self.kwargs)
+
+        return self
+
     async def fetch_experiences(self, query_type: QueryType):
         """Fetch a potentially perfect existing experience."""
 
-        req = self._get_req_identifier()
-        self._exps = await self.exp_manager.query_exps(req, query_type=query_type)
+        self._exps = await self.exp_manager.query_exps(self._req, query_type=query_type)
 
     def get_one_perfect_experience(self) -> Optional[Experience]:
         return self.exp_manager.extract_one_perfect_exp(self._exps)
@@ -107,26 +120,29 @@ class ExpCacheHandler(BaseModel):
     def save_experience(self):
         """Save the new experience."""
 
-        req = self._get_req_identifier()
-        exp = Experience(req=req, resp=self._result, metric=Metric(score=self._score))
+        exp = Experience(req=self._req, resp=self._result, metric=Metric(score=self._score))
 
         self.exp_manager.create_exp(exp)
 
-    def _get_req_identifier(self):
-        """Generate a unique request identifier based on the function and its arguments.
+    @classmethod
+    def generate_req_identifier(cls, func, *args, **kwargs) -> str:
+        """Generate a unique request identifier for any given function and its arguments.
 
-        Result Example:
-        - "write_prd-('2048',)-{}"
-        - "WritePRD.run-('2048',)-{}"
+        Serializing args and kwargs into JSON strings and replacing ',' with '~' and ':' with '!'.
+
+        Return Example:
+            SimpleClass.test_method@[1~2]@{"c"!3}
         """
-        if not self._req:
-            cls_name = get_class_name(self.func, *self.args)
-            func_name = f"{cls_name}.{self.func.__name__}" if cls_name else self.func.__name__
-            args = self.args[1:] if cls_name and len(self.args) >= 1 else self.args
+        cls_name = get_class_name(func)
+        func_name = f"{cls_name}.{func.__name__}" if cls_name else func.__name__
 
-            self._req = f"{func_name}-{args}-{self.kwargs}"
+        if cls_name and args and inspect.isfunction(func):
+            args = args[1:]
 
-        return self._req
+        args = cls._serialize_and_replace(args)
+        kwargs = cls._serialize_and_replace(kwargs)
+
+        return f"{func_name}@{args}@{kwargs}"
 
     @staticmethod
     def choose_wrapper(func, wrapped_func):
@@ -141,6 +157,11 @@ class ExpCacheHandler(BaseModel):
 
         return async_wrapper if asyncio.iscoroutinefunction(func) else sync_wrapper
 
+    @classmethod
+    def _serialize_and_replace(cls, data):
+        json_str = json.dumps(data)
+        return json_str.replace(", ", "~").replace(": ", "!")
+
     async def _execute_function(self):
         if self.pass_exps_to_func:
             return await self._execute_function_with_exps()
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 7382fe8f1..35ee5fdac 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -7,7 +7,7 @@ from pydantic import BaseModel, ConfigDict, model_validator
 from metagpt.config2 import Config, config
 from metagpt.exp_pool.schema import MAX_SCORE, Experience, QueryType
 from metagpt.rag.engines import SimpleEngine
-from metagpt.rag.schema import ChromaRetrieverConfig
+from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
 from metagpt.utils.exceptions import handle_exception
 
 
@@ -31,6 +31,7 @@ class ExperienceManager(BaseModel):
                 retriever_configs=[
                     ChromaRetrieverConfig(collection_name="experience_pool", persist_path=".chroma_exp_data")
                 ],
+                ranker_configs=[LLMRankerConfig()],
             )
         return self
 
diff --git a/metagpt/utils/reflection.py b/metagpt/utils/reflection.py
index 9b10a4b3e..fe852635f 100644
--- a/metagpt/utils/reflection.py
+++ b/metagpt/utils/reflection.py
@@ -19,24 +19,23 @@ def check_methods(C, *methods):
     return True
 
 
-def get_class_name(func, *args) -> str:
+def get_class_name(func) -> str:
     """Returns the class name of the object that a method belongs to.
 
-    - If `func` is a bound method, extracts the class name directly from the method.
-    - If `func` is an unbound method and `args` are provided, assumes the first argument is `self` and extracts the class name.
-    - Returns an empty string if neither condition is met.
+    - If `func` is a bound method or a class method, extracts the class name directly from the method.
+    - Returns an empty string if it's a regular function or cannot determine the class.
     """
     if inspect.ismethod(func):
+        if inspect.isclass(func.__self__):
+            return func.__self__.__name__
+
         return func.__self__.__class__.__name__
 
-    if inspect.isfunction(func) and "self" in inspect.signature(func).parameters and args:
-        return args[0].__class__.__name__
+    if inspect.isfunction(func):
+        qualname_parts = func.__qualname__.split(".")
+        if len(qualname_parts) > 1:
+            class_name = qualname_parts[-2]
+            if class_name.isidentifier():
+                return class_name
 
     return ""
-
-
-def get_func_or_method_name(func, *args) -> str:
-    """Function name, or method name with class name."""
-    cls_name = get_class_name(func, *args)
-
-    return f"{cls_name}.{func.__name__}" if cls_name else f"{func.__name__}"
diff --git a/tests/metagpt/exp_pool/test_decorator.py b/tests/metagpt/exp_pool/test_decorator.py
index 508229d18..bedc4e391 100644
--- a/tests/metagpt/exp_pool/test_decorator.py
+++ b/tests/metagpt/exp_pool/test_decorator.py
@@ -1,14 +1,28 @@
 import asyncio
+import inspect
 
 import pytest
 
-from metagpt.exp_pool.decorator import ExpCacheHandler
+from metagpt.exp_pool.decorator import ExpCacheHandler, exp_cache
 from metagpt.exp_pool.manager import ExperienceManager
 from metagpt.exp_pool.schema import Experience, QueryType, Score
 from metagpt.exp_pool.scorers import SimpleScorer
 from metagpt.rag.engines import SimpleEngine
 
 
+def for_test_function(a, b, c=None):
+    return a + b if c is None else a + b + c
+
+
+class ForTestClass:
+    def for_test_method(self, x, y):
+        return x * y
+
+    @classmethod
+    def for_test_class_method(cls, x, y):
+        return x**y
+
+
 class TestExpCache:
     @pytest.fixture
     def mock_func(self, mocker):
@@ -46,7 +60,7 @@ class TestExpCache:
         perfect_exp = Experience(req="req", resp="resp")
         mock_exp_manager.extract_one_perfect_exp.return_value = perfect_exp
 
-        # Execute
+        # Exec
         exp_cache_handler._exps = [perfect_exp]  # Simulate fetched experiences
         result = exp_cache_handler.get_one_perfect_experience()
 
@@ -60,7 +74,7 @@ class TestExpCache:
         mock_exp_manager.extract_one_perfect_exp.return_value = None
         mock_func.return_value = "Computed result"
 
-        # Execute
+        # Exec
         await exp_cache_handler.execute_function()
 
         # Assert
@@ -73,7 +87,7 @@ class TestExpCache:
         mock_scorer.evaluate.return_value = Score(value=100)
         exp_cache_handler._result = "Computed result"
 
-        # Execute
+        # Exec
         await exp_cache_handler.evaluate_experience()
         exp_cache_handler.save_experience()
 
@@ -84,12 +98,12 @@ class TestExpCache:
     @pytest.mark.asyncio
     async def test_async_function_execution_with_exps(self, exp_cache_handler, mock_exp_manager, mock_func):
         # Setup
-        exp_cache_handler.pass_exps = True
+        exp_cache_handler.pass_exps_to_func = True
         mock_func.return_value = "Async result with exps"
         mock_exp_manager.extract_one_perfect_exp.return_value = None
         exp_cache_handler._exps = [Experience(req="req", resp="resp")]
 
-        # Execute
+        # Exec
         await exp_cache_handler.execute_function()
 
         # Assert
@@ -99,11 +113,11 @@ class TestExpCache:
     def test_sync_function_execution_with_exps(self, mocker, exp_cache_handler, mock_exp_manager, mock_func):
         # Setup
         exp_cache_handler.func = mocker.Mock(return_value="Sync result with exps")
-        exp_cache_handler.pass_exps = True
+        exp_cache_handler.pass_exps_to_func = True
         mock_exp_manager.extract_one_perfect_exp.return_value = None
         exp_cache_handler._exps = [Experience(req="req", resp="resp")]
 
-        # Execute
+        # Exec
         asyncio.get_event_loop().run_until_complete(exp_cache_handler.execute_function())
 
         # Assert
@@ -114,7 +128,7 @@ class TestExpCache:
         # Setup
         mock_func = mocker.AsyncMock()
 
-        # Execute
+        # Exec
         wrapper = ExpCacheHandler.choose_wrapper(mock_func, exp_cache_handler.execute_function)
 
         # Assert
@@ -124,22 +138,80 @@ class TestExpCache:
         # Setup
         sync_func = mocker.Mock()
 
-        # Execute
+        # Exec
         wrapper = ExpCacheHandler.choose_wrapper(sync_func, exp_cache_handler.execute_function)
 
         # Assert
         assert not asyncio.iscoroutinefunction(wrapper), "Wrapper should be synchronous"
 
-    @pytest.mark.asyncio
-    async def test_generate_req_identifier(self, exp_cache_handler):
-        # Setup
-        exp_cache_handler.func = lambda x: x
-        exp_cache_handler.args = (42,)
-        exp_cache_handler.kwargs = {"y": 3.14}
+    @pytest.mark.parametrize(
+        "func, args, kwargs, expected",
+        [
+            (for_test_function, (1, 2), {"c": 3}, 'for_test_function@[1~2]@{"c"!3}'),
+            (ForTestClass().for_test_method, (4, 5), {}, "ForTestClass.for_test_method@[4~5]@{}"),
+            (ForTestClass.for_test_class_method, (6, 7), {}, "ForTestClass.for_test_class_method@[6~7]@{}"),
+            (for_test_function, (), {}, "for_test_function@[]@{}"),
+            (
+                for_test_function,
+                ("hello", [1, 2]),
+                {"key": "value"},
+                'for_test_function@["hello"~[1~2]]@{"key"!"value"}',
+            ),
+        ],
+    )
+    def test_generate_req_identifier(self, func, args, kwargs, expected):
+        req_identifier = ExpCacheHandler.generate_req_identifier(func, *args, **kwargs)
+        assert req_identifier == expected
 
-        # Execute
-        req_id = exp_cache_handler.generate_req_identifier()
+    @pytest.mark.asyncio
+    async def test_exp_cache_with_perfect_experience(self, mocker, mock_exp_manager):
+        # Mock perfect experience
+        perfect_exp = Experience(req="test_req", resp="perfect_response")
+        mock_exp_manager.query_exps = mocker.AsyncMock(return_value=[perfect_exp])
+        mock_exp_manager.extract_one_perfect_exp = mocker.MagicMock(return_value=perfect_exp)
+        async_mock_func = mocker.AsyncMock()
+
+        # Setup
+        decorated_func = exp_cache(async_mock_func, manager=mock_exp_manager)
+
+        # Exec
+        result: Experience = await decorated_func()
 
         # Assert
-        expected_id = "<lambda>_(42,)_{'y': 3.14}"
-        assert req_id == expected_id, "Request identifier should match the expected format"
+        assert result.resp == "perfect_response", "Should return the perfect experience response"
+        async_mock_func.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_exp_cache_without_perfect_experience(self, mocker, mock_exp_manager):
+        # Mock
+        mock_exp_manager.query_exps = mocker.AsyncMock(return_value=[])
+        mock_exp_manager.extract_one_perfect_exp = mocker.MagicMock(return_value=None)
+        async_mock_func = mocker.AsyncMock(return_value="computed_response")
+        async_mock_func.__signature__ = inspect.signature(for_test_function)
+
+        # Setup
+        decorated_func = exp_cache(async_mock_func, manager=mock_exp_manager)
+
+        # Exec
+        result = await decorated_func()
+
+        # Assert
+        assert result == "computed_response", "Should execute and return the function's response"
+        async_mock_func.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_exp_cache_saves_new_experience(self, mocker, mock_exp_manager, mock_scorer):
+        # Mock
+        mock_exp_manager.query_exps = mocker.AsyncMock(return_value=[])
+        mock_exp_manager.extract_one_perfect_exp = mocker.MagicMock(return_value=None)
+        async_mock_func = mocker.AsyncMock(return_value="computed_response")
+        mock_scorer.evaluate = mocker.AsyncMock(return_value=Score(value=100))
+
+        # Setup
+        decorated_func = exp_cache(async_mock_func, manager=mock_exp_manager, scorer=mock_scorer)
+
+        # Exec
+        await decorated_func()
+
+        # Assert
+        mock_exp_manager.create_exp.assert_called_once()
diff --git a/tests/metagpt/utils/test_reflection.py b/tests/metagpt/utils/test_reflection.py
index e78e1b400..58fd81619 100644
--- a/tests/metagpt/utils/test_reflection.py
+++ b/tests/metagpt/utils/test_reflection.py
@@ -1,29 +1,35 @@
-from metagpt.utils.reflection import get_func_or_method_name
+from metagpt.utils.reflection import get_class_name
 
 
-def simple_function():
-    pass
-
-
-class SampleClass:
-    def method(self):
+class SimpleFunction:
+    def function(self):
         pass
 
 
-class TestFunctionOrMethodName:
-    def test_simple_function(self):
-        assert get_func_or_method_name(simple_function) == "simple_function"
+class SampleClass:
+    @classmethod
+    def class_method(cls):
+        pass
 
-    def test_class_method_without_args(self):
-        sample_instance = SampleClass()
-        assert get_func_or_method_name(sample_instance.method) == "SampleClass.method"
+    def instance_method(self):
+        pass
 
-    def test_class_method_with_args(self):
-        sample_instance = SampleClass()
-        assert get_func_or_method_name(SampleClass.method, sample_instance) == "SampleClass.method"
 
-    def test_function_with_no_args(self):
-        assert get_func_or_method_name(simple_function) == "simple_function"
+def standalone_function():
+    pass
 
-    def test_method_without_instance(self):
-        assert get_func_or_method_name(SampleClass.method) == "method"
+
+class TestGetClassName:
+    def test_instance_method(self):
+        instance = SampleClass()
+        assert get_class_name(instance.instance_method) == "SampleClass"
+
+    def test_class_method(self):
+        assert get_class_name(SampleClass.class_method) == "SampleClass"
+
+    def test_standalone_function(self):
+        assert get_class_name(standalone_function) == ""
+
+    def test_function_within_simple_class(self):
+        instance = SimpleFunction()
+        assert get_class_name(instance.function) == "SimpleFunction"

From 0c4927f7246399bcb100fc91a63dbfae2793e430 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Wed, 12 Jun 2024 11:01:46 +0800
Subject: [PATCH 30/51] update comment

---
 metagpt/exp_pool/decorator.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 446220a47..2a3bf2fba 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -91,11 +91,12 @@ class ExpCacheHandler(BaseModel):
         return self
 
     async def fetch_experiences(self, query_type: QueryType):
-        """Fetch a potentially perfect existing experience."""
+        """Fetch experiences by query_type."""
 
         self._exps = await self.exp_manager.query_exps(self._req, query_type=query_type)
 
     def get_one_perfect_experience(self) -> Optional[Experience]:
+        """Get a potentially perfect experience."""
         return self.exp_manager.extract_one_perfect_exp(self._exps)
 
     async def execute_function(self):

From c624c0ffc74735e5467d6da3fa99adc0678f648c Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 8 Jul 2024 10:09:36 +0800
Subject: [PATCH 31/51] use llm cache to make exp_pool

---
 config/config2.example.yaml                   |   6 +-
 examples/exp_pool/decorator.py                |   7 +-
 examples/write_novel.py                       |   4 +-
 metagpt/actions/action.py                     |   2 +-
 metagpt/actions/action_node.py                |  31 +-
 metagpt/actions/design_api.py                 |   4 +-
 metagpt/actions/generate_questions.py         |   2 +-
 metagpt/actions/prepare_interview.py          |   2 +-
 metagpt/actions/project_management.py         |   4 +-
 metagpt/actions/write_code_an_draft.py        |   2 +-
 .../actions/write_code_plan_and_change_an.py  |   2 +-
 metagpt/actions/write_prd.py                  |   8 +-
 metagpt/actions/write_review.py               |   2 +-
 metagpt/configs/exp_pool_config.py            |   8 +-
 metagpt/exp_pool/context_builders/__init__.py |   7 +
 metagpt/exp_pool/context_builders/base.py     |  52 +++
 .../exp_pool/context_builders/role_zero.py    |  26 ++
 metagpt/exp_pool/context_builders/simple.py   |  24 ++
 metagpt/exp_pool/decorator.py                 | 165 +++++----
 metagpt/exp_pool/manager.py                   |  91 ++---
 metagpt/exp_pool/perfect_judges/__init__.py   |   6 +
 metagpt/exp_pool/perfect_judges/base.py       |  20 ++
 metagpt/exp_pool/perfect_judges/simple.py     |  27 ++
 metagpt/exp_pool/schema.py                    |   7 +-
 metagpt/exp_pool/scorers/__init__.py          |   6 +-
 metagpt/exp_pool/scorers/base.py              |   6 +-
 metagpt/exp_pool/scorers/simple.py            |   6 +-
 metagpt/roles/di/role_zero.py                 |  41 ++-
 metagpt/strategy/solver.py                    |   2 +-
 metagpt/utils/reflection.py                   |  23 --
 tests/metagpt/actions/test_action_node.py     |  14 +-
 tests/metagpt/actions/test_design_api_an.py   |   2 +-
 .../actions/test_project_management_an.py     |   4 +-
 tests/metagpt/actions/test_write_prd_an.py    |   2 +-
 .../test_base_context_builder.py              |  45 +++
 .../test_rolezero_context_builder.py          |  38 +++
 .../test_simple_context_builder.py            |  46 +++
 tests/metagpt/exp_pool/test_decorator.py      | 316 ++++++++----------
 tests/metagpt/exp_pool/test_manager.py        |  63 +++-
 .../test_simple_perfect_judge.py              |  40 +++
 .../test_scorers/test_simple_scorer.py        |  49 +++
 41 files changed, 844 insertions(+), 368 deletions(-)
 create mode 100644 metagpt/exp_pool/context_builders/__init__.py
 create mode 100644 metagpt/exp_pool/context_builders/base.py
 create mode 100644 metagpt/exp_pool/context_builders/role_zero.py
 create mode 100644 metagpt/exp_pool/context_builders/simple.py
 create mode 100644 metagpt/exp_pool/perfect_judges/__init__.py
 create mode 100644 metagpt/exp_pool/perfect_judges/base.py
 create mode 100644 metagpt/exp_pool/perfect_judges/simple.py
 create mode 100644 tests/metagpt/exp_pool/test_context_builders/test_base_context_builder.py
 create mode 100644 tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
 create mode 100644 tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py
 create mode 100644 tests/metagpt/exp_pool/test_perfect_judges/test_simple_perfect_judge.py
 create mode 100644 tests/metagpt/exp_pool/test_scorers/test_simple_scorer.py

diff --git a/config/config2.example.yaml b/config/config2.example.yaml
index c7b2cae2c..a3bd5c367 100644
--- a/config/config2.example.yaml
+++ b/config/config2.example.yaml
@@ -75,8 +75,10 @@ s3:
   bucket: "test"
 
 exp_pool:
-  enable_read: true
-  enable_write: true
+  enable_read: false 
+  enable_write: false
+  persist_path: .chroma_exp_data # The directory.
+  init_exp: false # If set to true, basic experiences associated with the roles will be added to the experience pool.
 
 azure_tts_subscription_key: "YOUR_SUBSCRIPTION_KEY"
 azure_tts_region: "eastus"
diff --git a/examples/exp_pool/decorator.py b/examples/exp_pool/decorator.py
index 3f6093e01..00726a0a8 100644
--- a/examples/exp_pool/decorator.py
+++ b/examples/exp_pool/decorator.py
@@ -7,16 +7,15 @@ from metagpt.exp_pool import exp_cache, exp_manager
 from metagpt.logs import logger
 
 
-@exp_cache(pass_exps_to_func=True)
-async def produce(req, exps=None):
-    logger.info(f"Previous experiences: {exps}")
+@exp_cache()
+async def produce(req=""):
     return f"{req} {uuid.uuid4().hex}"
 
 
 async def main():
     req = "Water"
 
-    resp = await produce(req)
+    resp = await produce(req=req)
     logger.info(f"The resp of `produce{req}` is: {resp}")
 
     exps = await exp_manager.query_exps(req)
diff --git a/examples/write_novel.py b/examples/write_novel.py
index a6e9ce05d..f49918fbb 100644
--- a/examples/write_novel.py
+++ b/examples/write_novel.py
@@ -50,9 +50,9 @@ async def generate_novel():
         "Fill the empty nodes with your own ideas. Be creative! Use your own words!"
         "I will tip you $100,000 if you write a good novel."
     )
-    novel_node = await ActionNode.from_pydantic(Novel).fill(context=instruction, llm=LLM())
+    novel_node = await ActionNode.from_pydantic(Novel).fill(req=instruction, llm=LLM())
     chap_node = await ActionNode.from_pydantic(Chapters).fill(
-        context=f"### instruction\n{instruction}\n### novel\n{novel_node.content}", llm=LLM()
+        req=f"### instruction\n{instruction}\n### novel\n{novel_node.content}", llm=LLM()
     )
     print(chap_node.instruct_content)
 
diff --git a/metagpt/actions/action.py b/metagpt/actions/action.py
index b760c96d8..8733947f5 100644
--- a/metagpt/actions/action.py
+++ b/metagpt/actions/action.py
@@ -90,7 +90,7 @@ class Action(SerializationMixin, ContextMixin, BaseModel):
         msgs = args[0]
         context = "## History Messages\n"
         context += "\n".join([f"{idx}: {i}" for idx, i in enumerate(reversed(msgs))])
-        return await self.node.fill(context=context, llm=self.llm)
+        return await self.node.fill(req=context, llm=self.llm)
 
     async def run(self, *args, **kwargs):
         """Run action"""
diff --git a/metagpt/actions/action_node.py b/metagpt/actions/action_node.py
index 48372f790..e1e0bddbb 100644
--- a/metagpt/actions/action_node.py
+++ b/metagpt/actions/action_node.py
@@ -18,6 +18,7 @@ from tenacity import retry, stop_after_attempt, wait_random_exponential
 
 from metagpt.actions.action_outcls_registry import register_action_outcls
 from metagpt.const import MARKDOWN_TITLE_PREFIX, USE_CONFIG_TIMEOUT
+from metagpt.exp_pool import exp_cache
 from metagpt.llm import BaseLLM
 from metagpt.logs import logger
 from metagpt.provider.postprocess.llm_output_postprocess import llm_output_postprocess
@@ -465,9 +466,33 @@ class ActionNode:
 
         return self
 
+    @classmethod
+    def deserialize_to_action_node(cls, serialized_data) -> "ActionNode":
+        """Customized deserialization, it will be triggered when a perfect experience is found.
+
+        ActionNode cannot be serialized, it throws an error 'cannot pickle 'SSLContext' object'.
+        """
+
+        class InstructContent:
+            def __init__(self, json_data):
+                self.json_data = json_data
+
+            def model_dump_json(self):
+                return self.json_data
+
+        action_node = cls(key="", expected_type=Type[str], instruction="", example="")
+        action_node.instruct_content = InstructContent(serialized_data)
+
+        return action_node
+
+    @exp_cache(
+        resp_serialize=lambda action_node: action_node.instruct_content.model_dump_json(),
+        resp_deserialize=lambda resp: ActionNode.deserialize_to_action_node(resp),
+    )
     async def fill(
         self,
-        context,
+        *,
+        req,
         llm,
         schema="json",
         mode="auto",
@@ -478,7 +503,7 @@ class ActionNode:
     ):
         """Fill the node(s) with mode.
 
-        :param context: Everything we should know when filling node.
+        :param req: Everything we should know when filling node.
         :param llm: Large Language Model with pre-defined system message.
         :param schema: json/markdown, determine example and output format.
          - raw: free form text
@@ -497,7 +522,7 @@ class ActionNode:
         :return: self
         """
         self.set_llm(llm)
-        self.set_context(context)
+        self.set_context(req)
         if self.schema:
             schema = self.schema
 
diff --git a/metagpt/actions/design_api.py b/metagpt/actions/design_api.py
index cc88171ff..1bfad20a2 100644
--- a/metagpt/actions/design_api.py
+++ b/metagpt/actions/design_api.py
@@ -178,12 +178,12 @@ class WriteDesign(Action):
         )
 
     async def _new_system_design(self, context):
-        node = await DESIGN_API_NODE.fill(context=context, llm=self.llm, schema=self.prompt_schema)
+        node = await DESIGN_API_NODE.fill(req=context, llm=self.llm, schema=self.prompt_schema)
         return node
 
     async def _merge(self, prd_doc, system_design_doc):
         context = NEW_REQ_TEMPLATE.format(old_design=system_design_doc.content, context=prd_doc.content)
-        node = await REFINED_DESIGN_NODE.fill(context=context, llm=self.llm, schema=self.prompt_schema)
+        node = await REFINED_DESIGN_NODE.fill(req=context, llm=self.llm, schema=self.prompt_schema)
         system_design_doc.content = node.instruct_content.model_dump_json()
         return system_design_doc
 
diff --git a/metagpt/actions/generate_questions.py b/metagpt/actions/generate_questions.py
index c96a37649..bf0ba6277 100644
--- a/metagpt/actions/generate_questions.py
+++ b/metagpt/actions/generate_questions.py
@@ -22,4 +22,4 @@ class GenerateQuestions(Action):
     name: str = "GenerateQuestions"
 
     async def run(self, context) -> ActionNode:
-        return await QUESTIONS.fill(context=context, llm=self.llm)
+        return await QUESTIONS.fill(req=context, llm=self.llm)
diff --git a/metagpt/actions/prepare_interview.py b/metagpt/actions/prepare_interview.py
index 04cc954d2..0a7eb6581 100644
--- a/metagpt/actions/prepare_interview.py
+++ b/metagpt/actions/prepare_interview.py
@@ -22,4 +22,4 @@ class PrepareInterview(Action):
     name: str = "PrepareInterview"
 
     async def run(self, context):
-        return await QUESTIONS.fill(context=context, llm=self.llm)
+        return await QUESTIONS.fill(req=context, llm=self.llm)
diff --git a/metagpt/actions/project_management.py b/metagpt/actions/project_management.py
index a39840bf1..ca2df2da9 100644
--- a/metagpt/actions/project_management.py
+++ b/metagpt/actions/project_management.py
@@ -151,12 +151,12 @@ class WriteTasks(Action):
         return task_doc
 
     async def _run_new_tasks(self, context: str):
-        node = await PM_NODE.fill(context, self.llm, schema=self.prompt_schema)
+        node = await PM_NODE.fill(req=context, llm=self.llm, schema=self.prompt_schema)
         return node
 
     async def _merge(self, system_design_doc, task_doc) -> Document:
         context = NEW_REQ_TEMPLATE.format(context=system_design_doc.content, old_task=task_doc.content)
-        node = await REFINED_PM_NODE.fill(context, self.llm, schema=self.prompt_schema)
+        node = await REFINED_PM_NODE.fill(req=context, llm=self.llm, schema=self.prompt_schema)
         task_doc.content = node.instruct_content.model_dump_json()
         return task_doc
 
diff --git a/metagpt/actions/write_code_an_draft.py b/metagpt/actions/write_code_an_draft.py
index ce030b0e9..4c3fd4c19 100644
--- a/metagpt/actions/write_code_an_draft.py
+++ b/metagpt/actions/write_code_an_draft.py
@@ -578,7 +578,7 @@ class WriteCodeAN(Action):
 
     async def run(self, context):
         self.llm.system_prompt = "You are an outstanding engineer and can implement any code"
-        return await WRITE_MOVE_NODE.fill(context=context, llm=self.llm, schema="json")
+        return await WRITE_MOVE_NODE.fill(req=context, llm=self.llm, schema="json")
 
 
 async def main():
diff --git a/metagpt/actions/write_code_plan_and_change_an.py b/metagpt/actions/write_code_plan_and_change_an.py
index 31482a94d..989df52f2 100644
--- a/metagpt/actions/write_code_plan_and_change_an.py
+++ b/metagpt/actions/write_code_plan_and_change_an.py
@@ -229,7 +229,7 @@ class WriteCodePlanAndChange(Action):
             code=await self.get_old_codes(),
         )
         logger.info("Writing code plan and change..")
-        return await WRITE_CODE_PLAN_AND_CHANGE_NODE.fill(context=context, llm=self.llm, schema="json")
+        return await WRITE_CODE_PLAN_AND_CHANGE_NODE.fill(req=context, llm=self.llm, schema="json")
 
     async def get_old_codes(self) -> str:
         old_codes = await self.repo.srcs.get_all()
diff --git a/metagpt/actions/write_prd.py b/metagpt/actions/write_prd.py
index 7199ec415..810823a24 100644
--- a/metagpt/actions/write_prd.py
+++ b/metagpt/actions/write_prd.py
@@ -211,7 +211,7 @@ class WritePRD(Action):
         context = CONTEXT_TEMPLATE.format(requirements=requirement, project_name=project_name)
         exclude = [PROJECT_NAME.key] if project_name else []
         node = await WRITE_PRD_NODE.fill(
-            context=context, llm=self.llm, exclude=exclude, schema=self.prompt_schema
+            req=context, llm=self.llm, exclude=exclude, schema=self.prompt_schema
         )  # schema=schema
         return node
 
@@ -238,7 +238,7 @@ class WritePRD(Action):
     async def _is_bugfix(self, context: str) -> bool:
         if not self.repo.code_files_exists():
             return False
-        node = await WP_ISSUE_TYPE_NODE.fill(context, self.llm)
+        node = await WP_ISSUE_TYPE_NODE.fill(req=context, llm=self.llm)
         return node.get("issue_type") == "BUG"
 
     async def get_related_docs(self, req: Document, docs: list[Document]) -> list[Document]:
@@ -248,14 +248,14 @@ class WritePRD(Action):
 
     async def _is_related(self, req: Document, old_prd: Document) -> bool:
         context = NEW_REQ_TEMPLATE.format(old_prd=old_prd.content, requirements=req.content)
-        node = await WP_IS_RELATIVE_NODE.fill(context, self.llm)
+        node = await WP_IS_RELATIVE_NODE.fill(req=context, llm=self.llm)
         return node.get("is_relative") == "YES"
 
     async def _merge(self, req: Document, related_doc: Document) -> Document:
         if not self.project_name:
             self.project_name = Path(self.project_path).name
         prompt = NEW_REQ_TEMPLATE.format(requirements=req.content, old_prd=related_doc.content)
-        node = await REFINED_PRD_NODE.fill(context=prompt, llm=self.llm, schema=self.prompt_schema)
+        node = await REFINED_PRD_NODE.fill(req=prompt, llm=self.llm, schema=self.prompt_schema)
         related_doc.content = node.instruct_content.model_dump_json()
         await self._rename_workspace(node)
         return related_doc
diff --git a/metagpt/actions/write_review.py b/metagpt/actions/write_review.py
index db8512946..907a1e990 100644
--- a/metagpt/actions/write_review.py
+++ b/metagpt/actions/write_review.py
@@ -36,4 +36,4 @@ class WriteReview(Action):
     name: str = "WriteReview"
 
     async def run(self, context):
-        return await WRITE_REVIEW_NODE.fill(context=context, llm=self.llm, schema="json")
+        return await WRITE_REVIEW_NODE.fill(req=context, llm=self.llm, schema="json")
diff --git a/metagpt/configs/exp_pool_config.py b/metagpt/configs/exp_pool_config.py
index 3f86173c1..0c92312da 100644
--- a/metagpt/configs/exp_pool_config.py
+++ b/metagpt/configs/exp_pool_config.py
@@ -4,5 +4,9 @@ from metagpt.utils.yaml_model import YamlModel
 
 
 class ExperiencePoolConfig(YamlModel):
-    enable_read: bool = Field(default=True, description="Enable to read from experience pool.")
-    enable_write: bool = Field(default=True, description="Enable to write to experience pool.")
+    enable_read: bool = Field(default=False, description="Enable to read from experience pool.")
+    enable_write: bool = Field(default=False, description="Enable to write to experience pool.")
+    persist_path: str = Field(default=".chroma_exp_data", description="The persist path for experience pool.")
+    init_exp: bool = Field(
+        default=False, description="Put some basic experiences associated with the roles into the experience pool."
+    )
diff --git a/metagpt/exp_pool/context_builders/__init__.py b/metagpt/exp_pool/context_builders/__init__.py
new file mode 100644
index 000000000..047558be0
--- /dev/null
+++ b/metagpt/exp_pool/context_builders/__init__.py
@@ -0,0 +1,7 @@
+"""Context builders init."""
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+from metagpt.exp_pool.context_builders.simple import SimpleContextBuilder
+from metagpt.exp_pool.context_builders.role_zero import RoleZeroContextBuilder
+
+__all__ = ["BaseContextBuilder", "SimpleContextBuilder", "RoleZeroContextBuilder"]
diff --git a/metagpt/exp_pool/context_builders/base.py b/metagpt/exp_pool/context_builders/base.py
new file mode 100644
index 000000000..e3fe320a6
--- /dev/null
+++ b/metagpt/exp_pool/context_builders/base.py
@@ -0,0 +1,52 @@
+"""Base context builder."""
+
+import re
+from abc import ABC, abstractmethod
+from typing import Any
+
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.exp_pool.schema import Experience
+
+EXP_TEMPLATE = """Given the request: {req}, We can get the response: {resp}, Which scored: {score}."""
+
+
+class BaseContextBuilder(BaseModel, ABC):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    exps: list[Experience] = []
+
+    @abstractmethod
+    async def build(self, *args, **kwargs) -> Any:
+        """Build context from parameters."""
+
+    def format_exps(self) -> str:
+        """Format experiences into a numbered list of strings."""
+
+        result = []
+        for i, exp in enumerate(self.exps, start=1):
+            result.append(f"{i}. " + EXP_TEMPLATE.format(req=exp.req, resp=exp.resp, score=exp.metric.score.val))
+
+        return "\n".join(result)
+
+    @staticmethod
+    def replace_content_between_markers(text: str, start_marker: str, end_marker: str, new_content: str) -> str:
+        """Replace the content between `start_marker` and `end_marker` in the text with `new_content`.
+
+        Args:
+            text (str): The original text.
+            new_content (str): The new content to replace the old content.
+            start_marker (str): The marker indicating the start of the content to be replaced, such as '# Example'.
+            end_marker (str): The marker indicating the end of the content to be replaced, such as '# Instruction'.
+
+        Returns:
+            str: The text with the content replaced.
+        """
+
+        pattern = re.compile(f"({start_marker}\n)(.*?)(\n{end_marker})", re.DOTALL)
+
+        def replacement(match):
+            return f"{match.group(1)}{new_content}\n{match.group(3)}"
+
+        replaced_text = pattern.sub(replacement, text)
+        return replaced_text
diff --git a/metagpt/exp_pool/context_builders/role_zero.py b/metagpt/exp_pool/context_builders/role_zero.py
new file mode 100644
index 000000000..60f71ef59
--- /dev/null
+++ b/metagpt/exp_pool/context_builders/role_zero.py
@@ -0,0 +1,26 @@
+"""RoleZero context builder."""
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+
+
+class RoleZeroContextBuilder(BaseContextBuilder):
+    async def build(self, *args, **kwargs) -> list[dict]:
+        """Builds the context by updating the req with formatted experiences.
+
+        If there are no experiences, retains the original examples in req, otherwise replaces the examples with the formatted experiences.
+        """
+
+        req = kwargs.get("req", [])
+        if not req:
+            return req
+
+        exps_str = self.format_exps()
+        if not exps_str:
+            return req
+
+        req[-1]["content"] = self.replace_example_content(req[-1].get("content", ""), exps_str)
+
+        return req
+
+    def replace_example_content(self, text: str, new_example_content: str) -> str:
+        return self.replace_content_between_markers(text, "# Example", "# Instruction", new_example_content)
diff --git a/metagpt/exp_pool/context_builders/simple.py b/metagpt/exp_pool/context_builders/simple.py
new file mode 100644
index 000000000..35e2e1c8a
--- /dev/null
+++ b/metagpt/exp_pool/context_builders/simple.py
@@ -0,0 +1,24 @@
+"""Simple context builder."""
+
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+
+SIMPLE_CONTEXT_TEMPLATE = """
+{req}
+
+### Experiences
+-----
+{exps}
+-----
+
+## Instruction
+Consider **Experiences** to generate a better answer.
+"""
+
+
+class SimpleContextBuilder(BaseContextBuilder):
+    async def build(self, *args, **kwargs) -> str:
+        req = kwargs.get("req", "")
+        exps = self.format_exps()
+
+        return SIMPLE_CONTEXT_TEMPLATE.format(req=req, exps=exps) if exps else req
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 2a3bf2fba..c518bb7ea 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -2,18 +2,19 @@
 
 import asyncio
 import functools
-import inspect
-import json
 from typing import Any, Callable, Optional, TypeVar
 
 from pydantic import BaseModel, ConfigDict, model_validator
 
+from metagpt.config2 import config
+from metagpt.exp_pool.context_builders import BaseContextBuilder, SimpleContextBuilder
 from metagpt.exp_pool.manager import ExperienceManager, exp_manager
+from metagpt.exp_pool.perfect_judges import BasePerfectJudge, SimplePerfectJudge
 from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
-from metagpt.exp_pool.scorers import ExperienceScorer, SimpleScorer
+from metagpt.exp_pool.scorers import BaseScorer, SimpleScorer
+from metagpt.logs import logger
 from metagpt.utils.async_helper import NestAsyncio
 from metagpt.utils.exceptions import handle_exception
-from metagpt.utils.reflection import get_class_name
 
 ReturnType = TypeVar("ReturnType")
 
@@ -21,42 +22,64 @@ ReturnType = TypeVar("ReturnType")
 def exp_cache(
     _func: Optional[Callable[..., ReturnType]] = None,
     query_type: QueryType = QueryType.SEMANTIC,
-    scorer: Optional[ExperienceScorer] = None,
     manager: Optional[ExperienceManager] = None,
-    pass_exps_to_func: bool = False,
+    scorer: Optional[BaseScorer] = None,
+    perfect_judge: Optional[BasePerfectJudge] = None,
+    context_builder: Optional[BaseContextBuilder] = None,
+    req_serialize: Optional[Callable[..., str]] = None,
+    resp_serialize: Optional[Callable[..., str]] = None,
+    resp_deserialize: Optional[Callable[[str], Any]] = None,
+    tag: Optional[str] = None,
 ):
     """Decorator to get a perfect experience, otherwise, it executes the function, and create a new experience.
 
-    This can be applied to both synchronous and asynchronous functions.
+    1. This can be applied to both synchronous and asynchronous functions.
+    2. The function must have a `req` parameter, and it must be provided as a keyword argument.
+    3. If `config.exp_pool.enable_read` is False, the decorator will just directly execute the function.
 
     Args:
         _func: Just to make the decorator more flexible, for example, it can be used directly with @exp_cache by default, without the need for @exp_cache().
         query_type: The type of query to be used when fetching experiences.
-        scorer: Evaluate experience. Default SimpleScorer.
-        manager: How to fetch, evaluate and save experience, etc. Default exp_manager.
-        pass_exps_to_func: To control whether imperfect experiences are passed to the function, if True, the func must have a parameter named 'exps'.
+        manager: How to fetch, evaluate and save experience, etc. Default to `exp_manager`.
+        scorer: Evaluate experience. Default to `SimpleScorer()`.
+        perfect_judge: Determines if an experience is perfect. Defaults to `SimplePerfectJudge()`.
+        context_builder: Build the context from exps and the function parameters. Default to `SimpleContextBuilder()`.
+        req_serialize: Serializes the request for storage. Defaults to `lambda req: str(req)`.
+        resp_serialize: Serializes the function's return value for storage. Defaults to `lambda resp: str(resp)`.
+        resp_deserialize: Deserializes the stored response back to the function's return value. Defaults to `lambda resp: resp`.
+        tag: An optional tag for the experience. Default to `ClassName.method_name` or `function_name`.
     """
 
     def decorator(func: Callable[..., ReturnType]) -> Callable[..., ReturnType]:
+        if not config.exp_pool.enable_read:
+            return func
+
         @functools.wraps(func)
         async def get_or_create(args: Any, kwargs: Any) -> ReturnType:
+            logger.info("exp_cache is enabled.")
             handler = ExpCacheHandler(
                 func=func,
                 args=args,
                 kwargs=kwargs,
+                query_type=query_type,
                 exp_manager=manager,
                 exp_scorer=scorer,
-                pass_exps_to_func=pass_exps_to_func,
+                exp_perfect_judge=perfect_judge,
+                context_builder=context_builder,
+                req_serialize=req_serialize,
+                resp_serialize=resp_serialize,
+                resp_deserialize=resp_deserialize,
+                tag=tag,
             )
 
-            await handler.fetch_experiences(query_type)
-            if exp := handler.get_one_perfect_experience():
+            await handler.fetch_experiences()
+            if exp := await handler.get_one_perfect_exp():
                 return exp
 
             await handler.execute_function()
             await handler.process_experience()
 
-            return handler._result
+            return handler._raw_resp
 
         return ExpCacheHandler.choose_wrapper(func, get_or_create)
 
@@ -69,39 +92,59 @@ class ExpCacheHandler(BaseModel):
     func: Callable
     args: Any
     kwargs: Any
+    query_type: QueryType = QueryType.SEMANTIC
     exp_manager: Optional[ExperienceManager] = None
-    exp_scorer: Optional[ExperienceScorer] = None
-    pass_exps_to_func: bool = False
+    exp_scorer: Optional[BaseScorer] = None
+    exp_perfect_judge: Optional[BasePerfectJudge] = None
+    context_builder: Optional[BaseContextBuilder] = None
+    req_serialize: Optional[Callable[..., str]] = None
+    resp_serialize: Optional[Callable[..., str]] = None
+    resp_deserialize: Optional[Callable[[str], Any]] = None
+    tag: Optional[str] = None
 
     _exps: list[Experience] = None
-    _result: Any = None
+    _req: str = ""
+    _resp: str = ""
+    _raw_resp: Any = None
     _score: Score = None
-    _req: str = None
 
     @model_validator(mode="after")
     def initialize(self):
-        if self.exp_manager is None:
-            self.exp_manager = exp_manager
+        self._validate_params()
 
-        if self.exp_scorer is None:
-            self.exp_scorer = SimpleScorer()
+        self.exp_manager = self.exp_manager or exp_manager
+        self.exp_scorer = self.exp_scorer or SimpleScorer()
+        self.exp_perfect_judge = self.exp_perfect_judge or SimplePerfectJudge()
+        self.context_builder = self.context_builder or SimpleContextBuilder()
+        self.req_serialize = self.req_serialize or (lambda resp: str(resp))
+        self.resp_serialize = self.resp_serialize or (lambda resp: str(resp))
+        self.resp_deserialize = self.resp_deserialize or (lambda resp: resp)
+        self.tag = self.tag or self._generate_tag()
 
-        self._req = self.generate_req_identifier(self.func, *self.args, **self.kwargs)
+        self._req = self.req_serialize(self.kwargs["req"])
 
         return self
 
-    async def fetch_experiences(self, query_type: QueryType):
+    async def fetch_experiences(self):
         """Fetch experiences by query_type."""
 
-        self._exps = await self.exp_manager.query_exps(self._req, query_type=query_type)
+        self._exps = await self.exp_manager.query_exps(self._req, query_type=self.query_type, tag=self.tag)
 
-    def get_one_perfect_experience(self) -> Optional[Experience]:
-        """Get a potentially perfect experience."""
-        return self.exp_manager.extract_one_perfect_exp(self._exps)
+    async def get_one_perfect_exp(self) -> Optional[Any]:
+        """Get a potentially perfect experience, and resolve resp."""
+
+        for exp in self._exps:
+            if await self.exp_perfect_judge.is_perfect_exp(exp, self._req, *self.args, **self.kwargs):
+                logger.info(f"Get one perfect experience: {exp.req[:20]}...")
+                return self.resp_deserialize(exp.resp)
+
+        return None
 
     async def execute_function(self):
-        """Execute the function, and save the result."""
-        self._result = await self._execute_function()
+        """Execute the function, and save resp."""
+
+        self._raw_resp = await self._execute_function()
+        self._resp = self.resp_serialize(self._raw_resp)
 
     @handle_exception
     async def process_experience(self):
@@ -110,41 +153,21 @@ class ExpCacheHandler(BaseModel):
         Evaluates and saves experience.
         Use `handle_exception` to ensure robustness, do not stop subsequent operations.
         """
+
         await self.evaluate_experience()
         self.save_experience()
 
     async def evaluate_experience(self):
         """Evaluate the experience, and save the score."""
 
-        self._score = await self.exp_scorer.evaluate(self.func, self._result, self.args, self.kwargs)
+        self._score = await self.exp_scorer.evaluate(self.func, self._resp, self.args, self.kwargs)
 
     def save_experience(self):
         """Save the new experience."""
 
-        exp = Experience(req=self._req, resp=self._result, metric=Metric(score=self._score))
-
+        exp = Experience(req=self._req, resp=self._resp, tag=self.tag, metric=Metric(score=self._score))
         self.exp_manager.create_exp(exp)
 
-    @classmethod
-    def generate_req_identifier(cls, func, *args, **kwargs) -> str:
-        """Generate a unique request identifier for any given function and its arguments.
-
-        Serializing args and kwargs into JSON strings and replacing ',' with '~' and ':' with '!'.
-
-        Return Example:
-            SimpleClass.test_method@[1~2]@{"c"!3}
-        """
-        cls_name = get_class_name(func)
-        func_name = f"{cls_name}.{func.__name__}" if cls_name else func.__name__
-
-        if cls_name and args and inspect.isfunction(func):
-            args = args[1:]
-
-        args = cls._serialize_and_replace(args)
-        kwargs = cls._serialize_and_replace(kwargs)
-
-        return f"{func_name}@{args}@{kwargs}"
-
     @staticmethod
     def choose_wrapper(func, wrapped_func):
         """Choose how to run wrapped_func based on whether the function is asynchronous."""
@@ -158,25 +181,31 @@ class ExpCacheHandler(BaseModel):
 
         return async_wrapper if asyncio.iscoroutinefunction(func) else sync_wrapper
 
-    @classmethod
-    def _serialize_and_replace(cls, data):
-        json_str = json.dumps(data)
-        return json_str.replace(", ", "~").replace(": ", "!")
+    def _validate_params(self):
+        if "req" not in self.kwargs:
+            raise ValueError("`req` must be provided as a keyword argument.")
+
+    def _generate_tag(self) -> str:
+        """Generates a tag for the self.func.
+
+        "ClassName.method_name" if the first argument is a class instance, otherwise just "function_name".
+        """
+
+        if self.args and hasattr(self.args[0], "__class__"):
+            cls_name = type(self.args[0]).__name__
+            return f"{cls_name}.{self.func.__name__}"
+
+        return self.func.__name__
+
+    async def _build_context(self) -> str:
+        self.context_builder.exps = self._exps
+
+        return await self.context_builder.build(*self.args, **self.kwargs)
 
     async def _execute_function(self):
-        if self.pass_exps_to_func:
-            return await self._execute_function_with_exps()
+        self.kwargs["req"] = await self._build_context()
 
-        return await self._execute_function_without_exps()
-
-    async def _execute_function_without_exps(self):
         if asyncio.iscoroutinefunction(self.func):
             return await self.func(*self.args, **self.kwargs)
 
         return self.func(*self.args, **self.kwargs)
-
-    async def _execute_function_with_exps(self):
-        if asyncio.iscoroutinefunction(self.func):
-            return await self.func(*self.args, **self.kwargs, exps=self._exps)
-
-        return self.func(*self.args, **self.kwargs, exps=self._exps)
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 35ee5fdac..276b1e8e3 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -1,13 +1,22 @@
 """Experience Manager."""
 
-from typing import Optional
-
+from llama_index.vector_stores.chroma import ChromaVectorStore
 from pydantic import BaseModel, ConfigDict, model_validator
 
 from metagpt.config2 import Config, config
-from metagpt.exp_pool.schema import MAX_SCORE, Experience, QueryType
+from metagpt.exp_pool.schema import (
+    DEFAULT_COLLECTION_NAME,
+    DEFAULT_SIMILARITY_TOP_K,
+    EntryType,
+    Experience,
+    Metric,
+    QueryType,
+    Score,
+)
+from metagpt.logs import logger
 from metagpt.rag.engines import SimpleEngine
 from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
+from metagpt.strategy.experience_retriever import ENGINEER_EXAMPLE, TL_EXAMPLE
 from metagpt.utils.exceptions import handle_exception
 
 
@@ -27,14 +36,33 @@ class ExperienceManager(BaseModel):
     @model_validator(mode="after")
     def initialize(self):
         if self.storage is None:
-            self.storage = SimpleEngine.from_objs(
-                retriever_configs=[
-                    ChromaRetrieverConfig(collection_name="experience_pool", persist_path=".chroma_exp_data")
-                ],
-                ranker_configs=[LLMRankerConfig()],
-            )
+            retriever_configs = [
+                ChromaRetrieverConfig(
+                    persist_path=self.config.exp_pool.persist_path,
+                    collection_name=DEFAULT_COLLECTION_NAME,
+                    similarity_top_k=DEFAULT_SIMILARITY_TOP_K,
+                )
+            ]
+            ranker_configs = [LLMRankerConfig()]
+
+            self.storage = SimpleEngine.from_objs(retriever_configs=retriever_configs, ranker_configs=ranker_configs)
+
+        self.init_exp_pool()
+
         return self
 
+    @handle_exception
+    def init_exp_pool(self):
+        if not self.config.exp_pool.init_exp:
+            return
+
+        if self._has_exps():
+            return
+
+        self._init_teamleader_exps()
+        self._init_engineer2_exps()
+        logger.info("`init_exp_pool` done.")
+
     @handle_exception
     def create_exp(self, exp: Experience):
         """Adds an experience to the storage if writing is enabled.
@@ -74,39 +102,26 @@ class ExperienceManager(BaseModel):
 
         return exps
 
-    def extract_one_perfect_exp(self, exps: list[Experience]) -> Optional[Experience]:
-        """Extracts the first 'perfect' experience from a list of experiences.
+    def _has_exps(self) -> bool:
+        vector_store: ChromaVectorStore = self.storage._retriever._vector_store
 
-        Args:
-            exps (list[Experience]): The experiences to evaluate.
+        return bool(vector_store._get(limit=1, where={}).ids)
 
-        Returns:
-            Optional[Experience]: The first perfect experience if found, otherwise None.
-        """
-        for exp in exps:
-            if self.is_perfect_exp(exp):
-                return exp
+    def _init_exp(self, req: str, resp: str, tag: str, metric: Metric = None):
+        exp = Experience(
+            req=req,
+            resp=resp,
+            entry_type=EntryType.MANUAL,
+            tag=tag,
+            metric=metric or Metric(score=Score(val=9, reason="Manual")),
+        )
+        self.create_exp(exp)
 
-        return None
+    def _init_teamleader_exps(self):
+        self._init_exp(req=TL_EXAMPLE, resp=TL_EXAMPLE, tag="TeamLeader.llm_cached_aask")
 
-    @staticmethod
-    def is_perfect_exp(exp: Experience) -> bool:
-        """Determines if an experience is considered 'perfect'.
-
-        Args:
-            exp (Experience): The experience to evaluate.
-
-        Returns:
-            bool: True if the experience is manually entered, otherwise False.
-        """
-        if not exp:
-            return False
-
-        # TODO: need more metrics
-        if exp.metric and exp.metric.score.val == MAX_SCORE:
-            return True
-
-        return False
+    def _init_engineer2_exps(self):
+        self._init_exp(req=ENGINEER_EXAMPLE, resp=ENGINEER_EXAMPLE, tag="Engineer2.llm_cached_aask")
 
 
 exp_manager = ExperienceManager()
diff --git a/metagpt/exp_pool/perfect_judges/__init__.py b/metagpt/exp_pool/perfect_judges/__init__.py
new file mode 100644
index 000000000..d8796c7c8
--- /dev/null
+++ b/metagpt/exp_pool/perfect_judges/__init__.py
@@ -0,0 +1,6 @@
+"""Perfect judges init."""
+
+from metagpt.exp_pool.perfect_judges.base import BasePerfectJudge
+from metagpt.exp_pool.perfect_judges.simple import SimplePerfectJudge
+
+__all__ = ["BasePerfectJudge", "SimplePerfectJudge"]
diff --git a/metagpt/exp_pool/perfect_judges/base.py b/metagpt/exp_pool/perfect_judges/base.py
new file mode 100644
index 000000000..293522993
--- /dev/null
+++ b/metagpt/exp_pool/perfect_judges/base.py
@@ -0,0 +1,20 @@
+"""Base perfect judge."""
+
+from abc import ABC, abstractmethod
+
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.exp_pool.schema import Experience
+
+
+class BasePerfectJudge(BaseModel, ABC):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    @abstractmethod
+    async def is_perfect_exp(self, exp: Experience, serialized_req: str, *args, **kwargs) -> bool:
+        """Determine whether the experience is perfect.
+
+        Args:
+            exp (Experience): The experience to evaluate.
+            serialized_req (str): The serialized request to compare against the experience's request.
+        """
diff --git a/metagpt/exp_pool/perfect_judges/simple.py b/metagpt/exp_pool/perfect_judges/simple.py
new file mode 100644
index 000000000..37ede95c3
--- /dev/null
+++ b/metagpt/exp_pool/perfect_judges/simple.py
@@ -0,0 +1,27 @@
+"""Simple perfect judge."""
+
+
+from pydantic import ConfigDict
+
+from metagpt.exp_pool.perfect_judges.base import BasePerfectJudge
+from metagpt.exp_pool.schema import MAX_SCORE, Experience
+
+
+class SimplePerfectJudge(BasePerfectJudge):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    async def is_perfect_exp(self, exp: Experience, serialized_req: str, *args, **kwargs) -> bool:
+        """Determine whether the experience is perfect.
+
+        Args:
+            exp (Experience): The experience to evaluate.
+            serialized_req (str): The serialized request to compare against the experience's request.
+
+        Returns:
+            bool: True if the serialized request matches the experience's request and the experience's score is perfect, False otherwise.
+        """
+
+        if not exp.metric or not exp.metric.score:
+            return False
+
+        return serialized_req == exp.req and exp.metric.score.val == MAX_SCORE
diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index 9fc665cca..d59478742 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -1,13 +1,16 @@
 """Experience schema."""
 
 from enum import Enum
-from typing import Any, Optional
+from typing import Optional
 
 from llama_index.core.schema import TextNode
 from pydantic import BaseModel, Field
 
 MAX_SCORE = 10
 
+DEFAULT_COLLECTION_NAME = "experience_pool"
+DEFAULT_SIMILARITY_TOP_K = 2
+
 
 class QueryType(str, Enum):
     """Type of query experiences."""
@@ -59,7 +62,7 @@ class Experience(BaseModel):
     """Experience."""
 
     req: str = Field(..., description="")
-    resp: Any = Field(..., description="The type is string/json/code.")
+    resp: str = Field(..., description="The type is string/json/code.")
     metric: Optional[Metric] = Field(default=None, description="Metric.")
     exp_type: ExperienceType = Field(default=ExperienceType.SUCCESS, description="The type of experience.")
     entry_type: EntryType = Field(default=EntryType.AUTOMATIC, description="Type of entry: Manual or Automatic.")
diff --git a/metagpt/exp_pool/scorers/__init__.py b/metagpt/exp_pool/scorers/__init__.py
index 85bea88ff..caa845b14 100644
--- a/metagpt/exp_pool/scorers/__init__.py
+++ b/metagpt/exp_pool/scorers/__init__.py
@@ -1,6 +1,6 @@
-"""Experience scorers init."""
+"""Scorers init."""
 
-from metagpt.exp_pool.scorers.base import ExperienceScorer
+from metagpt.exp_pool.scorers.base import BaseScorer
 from metagpt.exp_pool.scorers.simple import SimpleScorer
 
-__all__ = ["ExperienceScorer", "SimpleScorer"]
+__all__ = ["BaseScorer", "SimpleScorer"]
diff --git a/metagpt/exp_pool/scorers/base.py b/metagpt/exp_pool/scorers/base.py
index a9d30cffe..94623c30f 100644
--- a/metagpt/exp_pool/scorers/base.py
+++ b/metagpt/exp_pool/scorers/base.py
@@ -1,6 +1,6 @@
-"""Experience Scorers."""
+"""Base scorer."""
 
-from abc import abstractmethod
+from abc import ABC, abstractmethod
 from typing import Any, Callable
 
 from pydantic import BaseModel, ConfigDict
@@ -8,7 +8,7 @@ from pydantic import BaseModel, ConfigDict
 from metagpt.exp_pool.schema import Score
 
 
-class ExperienceScorer(BaseModel):
+class BaseScorer(BaseModel, ABC):
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
     @abstractmethod
diff --git a/metagpt/exp_pool/scorers/simple.py b/metagpt/exp_pool/scorers/simple.py
index 84995b60f..1fda189d1 100644
--- a/metagpt/exp_pool/scorers/simple.py
+++ b/metagpt/exp_pool/scorers/simple.py
@@ -1,4 +1,4 @@
-"""Simple Scorer."""
+"""Simple scorer."""
 
 import inspect
 import json
@@ -7,7 +7,7 @@ from typing import Any, Callable
 from pydantic import Field
 
 from metagpt.exp_pool.schema import Score
-from metagpt.exp_pool.scorers.base import ExperienceScorer
+from metagpt.exp_pool.scorers.base import BaseScorer
 from metagpt.llm import LLM
 from metagpt.provider.base_llm import BaseLLM
 from metagpt.utils.common import CodeParser
@@ -54,7 +54,7 @@ Follow instructions, generate output and make sure it follows the **Constraint**
 """
 
 
-class SimpleScorer(ExperienceScorer):
+class SimpleScorer(BaseScorer):
     llm: BaseLLM = Field(default_factory=LLM)
 
     async def evaluate(self, func: Callable, result: Any, args: tuple = None, kwargs: dict = None) -> Score:
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 906c5583c..e2a4cec78 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import copy
 import inspect
 import json
 import re
@@ -10,8 +11,14 @@ from pydantic import model_validator
 
 from metagpt.actions import Action
 from metagpt.actions.di.run_command import RunCommand
+from metagpt.exp_pool import exp_cache
+from metagpt.exp_pool.context_builders import RoleZeroContextBuilder
 from metagpt.logs import logger
-from metagpt.prompts.di.role_zero import CMD_PROMPT, ROLE_INSTRUCTION, JSON_REPAIR_PROMPT
+from metagpt.prompts.di.role_zero import (
+    CMD_PROMPT,
+    JSON_REPAIR_PROMPT,
+    ROLE_INSTRUCTION,
+)
 from metagpt.roles import Role
 from metagpt.schema import AIMessage, Message, UserMessage
 from metagpt.strategy.experience_retriever import DummyExpRetriever, ExpRetriever
@@ -21,8 +28,8 @@ from metagpt.tools.libs.editor import Editor
 from metagpt.tools.tool_recommend import BM25ToolRecommender, ToolRecommender
 from metagpt.tools.tool_registry import register_tool
 from metagpt.utils.common import CodeParser
+from metagpt.utils.repair_llm_raw_output import RepairType, repair_llm_raw_output
 from metagpt.utils.report import ThoughtReporter
-from metagpt.utils.repair_llm_raw_output import repair_llm_raw_output, RepairType
 
 
 @register_tool(include_functions=["ask_human", "reply_to_human"])
@@ -154,11 +161,37 @@ class RoleZero(Role):
         context = self.llm.format_msg(memory + [UserMessage(content=prompt)])
         # print(*context, sep="\n" + "*" * 5 + "\n")
         async with ThoughtReporter(enable_llm_stream=True):
-            self.command_rsp = await self.llm.aask(context, system_msgs=self.system_msg)
+            self.command_rsp = await self.llm_cached_aask(req=context, system_msgs=self.system_msg)
         self.rc.memory.add(AIMessage(content=self.command_rsp))
 
         return True
 
+    @exp_cache(context_builder=RoleZeroContextBuilder(), req_serialize=lambda req: RoleZero._req_serialize(req))
+    async def llm_cached_aask(self, *, req: list[dict], system_msgs: list[str]) -> str:
+        return await self.llm.aask(req, system_msgs=system_msgs)
+
+    @staticmethod
+    def _req_serialize(req: list[dict]) -> str:
+        """Serialize the request for database storage, ensuring it is a string.
+
+        This function deep copies the request and modifies the content of the last element
+        to remove unnecessary sections, making the request more concise.
+        """
+
+        req_copy = copy.deepcopy(req)
+
+        last_content = req_copy[-1]["content"]
+        last_content = RoleZeroContextBuilder.replace_content_between_markers(
+            last_content, "# Data Structure", "# Current Plan", ""
+        )
+        last_content = RoleZeroContextBuilder.replace_content_between_markers(
+            last_content, "# Example", "# Instruction", ""
+        )
+
+        req_copy[-1]["content"] = last_content
+
+        return json.dumps(req_copy)
+
     async def _act(self) -> Message:
         if self.use_fixed_sop:
             return await super()._act()
@@ -166,7 +199,7 @@ class RoleZero(Role):
         try:
             commands = CodeParser.parse_code(block=None, lang="json", text=self.command_rsp)
             commands = json.loads(repair_llm_raw_output(output=commands, req_keys=[None], repair_type=RepairType.JSON))
-        except json.JSONDecodeError as e:
+        except json.JSONDecodeError:
             commands = await self.llm.aask(msg=JSON_REPAIR_PROMPT.format(json_data=self.command_rsp))
             commands = json.loads(CodeParser.parse_code(block=None, lang="json", text=commands))
         except Exception as e:
diff --git a/metagpt/strategy/solver.py b/metagpt/strategy/solver.py
index e532f736b..4aedb42aa 100644
--- a/metagpt/strategy/solver.py
+++ b/metagpt/strategy/solver.py
@@ -39,7 +39,7 @@ class NaiveSolver(BaseSolver):
         self.graph.topological_sort()
         for key in self.graph.execution_order:
             op = self.graph.nodes[key]
-            await op.fill(self.context, self.llm, mode="root")
+            await op.fill(req=self.context, llm=self.llm, mode="root")
 
 
 class TOTSolver(BaseSolver):
diff --git a/metagpt/utils/reflection.py b/metagpt/utils/reflection.py
index fe852635f..8b8237ae7 100644
--- a/metagpt/utils/reflection.py
+++ b/metagpt/utils/reflection.py
@@ -1,5 +1,4 @@
 """class tools, including method inspection, class attributes, inheritance relationships, etc."""
-import inspect
 
 
 def check_methods(C, *methods):
@@ -17,25 +16,3 @@ def check_methods(C, *methods):
         else:
             return NotImplemented
     return True
-
-
-def get_class_name(func) -> str:
-    """Returns the class name of the object that a method belongs to.
-
-    - If `func` is a bound method or a class method, extracts the class name directly from the method.
-    - Returns an empty string if it's a regular function or cannot determine the class.
-    """
-    if inspect.ismethod(func):
-        if inspect.isclass(func.__self__):
-            return func.__self__.__name__
-
-        return func.__self__.__class__.__name__
-
-    if inspect.isfunction(func):
-        qualname_parts = func.__qualname__.split(".")
-        if len(qualname_parts) > 1:
-            class_name = qualname_parts[-2]
-            if class_name.isidentifier():
-                return class_name
-
-    return ""
diff --git a/tests/metagpt/actions/test_action_node.py b/tests/metagpt/actions/test_action_node.py
index bc85925a8..23779c984 100644
--- a/tests/metagpt/actions/test_action_node.py
+++ b/tests/metagpt/actions/test_action_node.py
@@ -91,10 +91,10 @@ async def test_action_node_two_layer():
     assert node_b in root.children.values()
 
     # FIXME: ADD MARKDOWN SUPPORT. NEED TO TUNE MARKDOWN SYMBOL FIRST.
-    answer1 = await root.fill(context="what's the answer to 123+456?", schema="json", strgy="simple", llm=LLM())
+    answer1 = await root.fill(req="what's the answer to 123+456?", schema="json", strgy="simple", llm=LLM())
     assert "579" in answer1.content
 
-    answer2 = await root.fill(context="what's the answer to 123+456?", schema="json", strgy="complex", llm=LLM())
+    answer2 = await root.fill(req="what's the answer to 123+456?", schema="json", strgy="complex", llm=LLM())
     assert "579" in answer2.content
 
 
@@ -112,7 +112,7 @@ async def test_action_node_review():
     with pytest.raises(RuntimeError):
         _ = await node_a.review()
 
-    _ = await node_a.fill(context=None, llm=LLM())
+    _ = await node_a.fill(req=None, llm=LLM())
     setattr(node_a.instruct_content, key, "game snake")  # wrong content to review
 
     review_comments = await node_a.review(review_mode=ReviewMode.AUTO)
@@ -126,7 +126,7 @@ async def test_action_node_review():
     with pytest.raises(RuntimeError):
         _ = await node.review()
 
-    _ = await node.fill(context=None, llm=LLM())
+    _ = await node.fill(req=None, llm=LLM())
 
     review_comments = await node.review(review_mode=ReviewMode.AUTO)
     assert len(review_comments) == 1
@@ -151,7 +151,7 @@ async def test_action_node_revise():
     with pytest.raises(RuntimeError):
         _ = await node_a.review()
 
-    _ = await node_a.fill(context=None, llm=LLM())
+    _ = await node_a.fill(req=None, llm=LLM())
     setattr(node_a.instruct_content, key, "game snake")  # wrong content to revise
     revise_contents = await node_a.revise(revise_mode=ReviseMode.AUTO)
     assert len(revise_contents) == 1
@@ -164,7 +164,7 @@ async def test_action_node_revise():
     with pytest.raises(RuntimeError):
         _ = await node.revise()
 
-    _ = await node.fill(context=None, llm=LLM())
+    _ = await node.fill(req=None, llm=LLM())
     setattr(node.instruct_content, key, "game snake")
     revise_contents = await node.revise(revise_mode=ReviseMode.AUTO)
     assert len(revise_contents) == 1
@@ -257,7 +257,7 @@ async def test_action_node_with_image(mocker):
     invoice_path = Path(__file__).parent.joinpath("..", "..", "data", "invoices", "invoice-2.png")
     img_base64 = encode_image(invoice_path)
     mocker.patch("metagpt.provider.openai_api.OpenAILLM._cons_kwargs", _cons_kwargs)
-    node = await invoice.fill(context="", llm=LLM(), images=[img_base64])
+    node = await invoice.fill(req="", llm=LLM(), images=[img_base64])
     assert node.instruct_content.invoice
 
 
diff --git a/tests/metagpt/actions/test_design_api_an.py b/tests/metagpt/actions/test_design_api_an.py
index 3d11f200d..4ed3cb362 100644
--- a/tests/metagpt/actions/test_design_api_an.py
+++ b/tests/metagpt/actions/test_design_api_an.py
@@ -38,7 +38,7 @@ async def test_write_design_an(mocker):
     mocker.patch("metagpt.actions.design_api_an.REFINED_DESIGN_NODE.fill", return_value=root)
 
     prompt = NEW_REQ_TEMPLATE.format(old_design=DESIGN_SAMPLE, context=dict_to_markdown(REFINED_PRD_JSON))
-    node = await REFINED_DESIGN_NODE.fill(prompt, llm)
+    node = await REFINED_DESIGN_NODE.fill(req=prompt, llm=llm)
 
     assert "Refined Implementation Approach" in node.instruct_content.model_dump()
     assert "Refined File list" in node.instruct_content.model_dump()
diff --git a/tests/metagpt/actions/test_project_management_an.py b/tests/metagpt/actions/test_project_management_an.py
index 5a65e50c9..6d41109c9 100644
--- a/tests/metagpt/actions/test_project_management_an.py
+++ b/tests/metagpt/actions/test_project_management_an.py
@@ -42,7 +42,7 @@ async def test_project_management_an(mocker):
     root.instruct_content.model_dump = mock_task_json
     mocker.patch("metagpt.actions.project_management_an.PM_NODE.fill", return_value=root)
 
-    node = await PM_NODE.fill(dict_to_markdown(REFINED_DESIGN_JSON), llm)
+    node = await PM_NODE.fill(req=dict_to_markdown(REFINED_DESIGN_JSON), llm=llm)
 
     assert "Logic Analysis" in node.instruct_content.model_dump()
     assert "Task list" in node.instruct_content.model_dump()
@@ -59,7 +59,7 @@ async def test_project_management_an_inc(mocker):
     mocker.patch("metagpt.actions.project_management_an.REFINED_PM_NODE.fill", return_value=root)
 
     prompt = NEW_REQ_TEMPLATE.format(old_task=TASK_SAMPLE, context=dict_to_markdown(REFINED_DESIGN_JSON))
-    node = await REFINED_PM_NODE.fill(prompt, llm)
+    node = await REFINED_PM_NODE.fill(req=prompt, llm=llm)
 
     assert "Refined Logic Analysis" in node.instruct_content.model_dump()
     assert "Refined Task list" in node.instruct_content.model_dump()
diff --git a/tests/metagpt/actions/test_write_prd_an.py b/tests/metagpt/actions/test_write_prd_an.py
index 378ce42c3..b6e92d3d6 100644
--- a/tests/metagpt/actions/test_write_prd_an.py
+++ b/tests/metagpt/actions/test_write_prd_an.py
@@ -39,7 +39,7 @@ async def test_write_prd_an(mocker):
         requirements=NEW_REQUIREMENT_SAMPLE,
         old_prd=PRD_SAMPLE,
     )
-    node = await REFINED_PRD_NODE.fill(prompt, llm)
+    node = await REFINED_PRD_NODE.fill(req=prompt, llm=llm)
 
     assert "Refined Requirements" in node.instruct_content.model_dump()
     assert "Refined Product Goals" in node.instruct_content.model_dump()
diff --git a/tests/metagpt/exp_pool/test_context_builders/test_base_context_builder.py b/tests/metagpt/exp_pool/test_context_builders/test_base_context_builder.py
new file mode 100644
index 000000000..17696e1b4
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_context_builders/test_base_context_builder.py
@@ -0,0 +1,45 @@
+import pytest
+
+from metagpt.exp_pool.context_builders.base import (
+    EXP_TEMPLATE,
+    BaseContextBuilder,
+    Experience,
+)
+from metagpt.exp_pool.schema import Metric, Score
+
+
+class TestBaseContextBuilder:
+    class ConcreteContextBuilder(BaseContextBuilder):
+        async def build(self, *args, **kwargs):
+            pass
+
+    @pytest.fixture
+    def context_builder(self):
+        return self.ConcreteContextBuilder()
+
+    def test_format_exps(self, context_builder):
+        exp1 = Experience(req="req1", resp="resp1", metric=Metric(score=Score(val=8)))
+        exp2 = Experience(req="req2", resp="resp2", metric=Metric(score=Score(val=9)))
+        context_builder.exps = [exp1, exp2]
+
+        result = context_builder.format_exps()
+        expected = "\n".join(
+            [
+                f"1. {EXP_TEMPLATE.format(req='req1', resp='resp1', score=8)}",
+                f"2. {EXP_TEMPLATE.format(req='req2', resp='resp2', score=9)}",
+            ]
+        )
+        assert result == expected
+
+    def test_replace_content_between_markers(self):
+        text = "Start\n# Example\nOld content\n# Instruction\nEnd"
+        new_content = "New content"
+        result = BaseContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
+        expected = "Start\n# Example\nNew content\n\n# Instruction\nEnd"
+        assert result == expected
+
+    def test_replace_content_between_markers_no_match(self):
+        text = "Start\nNo markers\nEnd"
+        new_content = "New content"
+        result = BaseContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
+        assert result == text
diff --git a/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py b/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
new file mode 100644
index 000000000..0ea04432d
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
@@ -0,0 +1,38 @@
+import pytest
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+from metagpt.exp_pool.context_builders.role_zero import RoleZeroContextBuilder
+
+
+class TestRoleZeroContextBuilder:
+    @pytest.fixture
+    def context_builder(self):
+        return RoleZeroContextBuilder()
+
+    @pytest.mark.asyncio
+    async def test_build_empty_req(self, context_builder):
+        result = await context_builder.build(req=[])
+        assert result == []
+
+    @pytest.mark.asyncio
+    async def test_build_no_experiences(self, context_builder, mocker):
+        mocker.patch.object(BaseContextBuilder, "format_exps", return_value="")
+        req = [{"content": "Original content"}]
+        result = await context_builder.build(req=req)
+        assert result == req
+
+    @pytest.mark.asyncio
+    async def test_build_with_experiences(self, context_builder, mocker):
+        mocker.patch.object(BaseContextBuilder, "format_exps", return_value="Formatted experiences")
+        mocker.patch.object(RoleZeroContextBuilder, "replace_example_content", return_value="Updated content")
+        req = [{"content": "Original content"}]
+        result = await context_builder.build(req=req)
+        assert result == [{"content": "Updated content"}]
+
+    def test_replace_example_content(self, context_builder, mocker):
+        mocker.patch.object(BaseContextBuilder, "replace_content_between_markers", return_value="Replaced content")
+        result = context_builder.replace_example_content("Original text", "New example content")
+        assert result == "Replaced content"
+        context_builder.replace_content_between_markers.assert_called_once_with(
+            "Original text", "# Example", "# Instruction", "New example content"
+        )
diff --git a/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py b/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py
new file mode 100644
index 000000000..e96addab9
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py
@@ -0,0 +1,46 @@
+import pytest
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+from metagpt.exp_pool.context_builders.simple import (
+    SIMPLE_CONTEXT_TEMPLATE,
+    SimpleContextBuilder,
+)
+
+
+class TestSimpleContextBuilder:
+    @pytest.fixture
+    def context_builder(self):
+        return SimpleContextBuilder()
+
+    @pytest.mark.asyncio
+    async def test_build_with_experiences(self, context_builder, mocker):
+        # Mock the format_exps method
+        mock_exps = "Mocked experiences"
+        mocker.patch.object(BaseContextBuilder, "format_exps", return_value=mock_exps)
+
+        req = "Test request"
+        result = await context_builder.build(req=req)
+
+        expected = SIMPLE_CONTEXT_TEMPLATE.format(req=req, exps=mock_exps)
+        assert result == expected
+
+    @pytest.mark.asyncio
+    async def test_build_without_experiences(self, context_builder, mocker):
+        # Mock the format_exps method to return an empty string
+        mocker.patch.object(BaseContextBuilder, "format_exps", return_value="")
+
+        req = "Test request"
+        result = await context_builder.build(req=req)
+
+        assert result == req
+
+    @pytest.mark.asyncio
+    async def test_build_without_req(self, context_builder, mocker):
+        # Mock the format_exps method
+        mock_exps = "Mocked experiences"
+        mocker.patch.object(BaseContextBuilder, "format_exps", return_value=mock_exps)
+
+        result = await context_builder.build()
+
+        expected = SIMPLE_CONTEXT_TEMPLATE.format(req="", exps=mock_exps)
+        assert result == expected
diff --git a/tests/metagpt/exp_pool/test_decorator.py b/tests/metagpt/exp_pool/test_decorator.py
index bedc4e391..c0b3fe36d 100644
--- a/tests/metagpt/exp_pool/test_decorator.py
+++ b/tests/metagpt/exp_pool/test_decorator.py
@@ -1,29 +1,17 @@
 import asyncio
-import inspect
 
 import pytest
 
+from metagpt.exp_pool.context_builders import SimpleContextBuilder
 from metagpt.exp_pool.decorator import ExpCacheHandler, exp_cache
 from metagpt.exp_pool.manager import ExperienceManager
+from metagpt.exp_pool.perfect_judges import SimplePerfectJudge
 from metagpt.exp_pool.schema import Experience, QueryType, Score
 from metagpt.exp_pool.scorers import SimpleScorer
 from metagpt.rag.engines import SimpleEngine
 
 
-def for_test_function(a, b, c=None):
-    return a + b if c is None else a + b + c
-
-
-class ForTestClass:
-    def for_test_method(self, x, y):
-        return x * y
-
-    @classmethod
-    def for_test_class_method(cls, x, y):
-        return x**y
-
-
-class TestExpCache:
+class TestExpCacheHandler:
     @pytest.fixture
     def mock_func(self, mocker):
         return mocker.AsyncMock()
@@ -34,7 +22,6 @@ class TestExpCache:
         manager.storage = mocker.MagicMock(spec=SimpleEngine)
         manager.query_exps = mocker.AsyncMock()
         manager.create_exp = mocker.MagicMock()
-        manager.extract_one_perfect_exp = mocker.MagicMock()
         return manager
 
     @pytest.fixture
@@ -44,174 +31,165 @@ class TestExpCache:
         return scorer
 
     @pytest.fixture
-    def exp_cache_handler(self, mock_func, mock_exp_manager, mock_scorer):
+    def mock_perfect_judge(self, mocker):
+        return mocker.MagicMock(spec=SimplePerfectJudge)
+
+    @pytest.fixture
+    def mock_context_builder(self, mocker):
+        return mocker.MagicMock(spec=SimpleContextBuilder)
+
+    @pytest.fixture
+    def exp_cache_handler(self, mock_func, mock_exp_manager, mock_scorer, mock_perfect_judge, mock_context_builder):
         return ExpCacheHandler(
-            func=mock_func, args=(), kwargs={}, exp_manager=mock_exp_manager, exp_scorer=mock_scorer, pass_exps=False
+            func=mock_func,
+            args=(),
+            kwargs={"req": "test_req"},
+            exp_manager=mock_exp_manager,
+            exp_scorer=mock_scorer,
+            exp_perfect_judge=mock_perfect_judge,
+            context_builder=mock_context_builder,
         )
 
     @pytest.mark.asyncio
     async def test_fetch_experiences(self, exp_cache_handler, mock_exp_manager):
-        await exp_cache_handler.fetch_experiences(QueryType.SEMANTIC)
-        mock_exp_manager.query_exps.assert_called_once()
+        mock_exp_manager.query_exps.return_value = [Experience(req="test_req", resp="test_resp")]
+        await exp_cache_handler.fetch_experiences()
+        mock_exp_manager.query_exps.assert_called_once_with(
+            "test_req", query_type=QueryType.SEMANTIC, tag=exp_cache_handler.tag
+        )
+        assert len(exp_cache_handler._exps) == 1
 
     @pytest.mark.asyncio
-    async def test_perfect_experience_found(self, exp_cache_handler, mock_exp_manager, mock_func):
-        # Setup: Assume perfect experience is found
-        perfect_exp = Experience(req="req", resp="resp")
-        mock_exp_manager.extract_one_perfect_exp.return_value = perfect_exp
-
-        # Exec
-        exp_cache_handler._exps = [perfect_exp]  # Simulate fetched experiences
-        result = exp_cache_handler.get_one_perfect_experience()
-
-        # Assert
-        assert result.resp == "resp"
-        mock_func.assert_not_called()  # Function should not be called
+    async def test_get_one_perfect_exp(self, exp_cache_handler, mock_perfect_judge):
+        exp = Experience(req="test_req", resp="perfect_resp")
+        exp_cache_handler._exps = [exp]
+        mock_perfect_judge.is_perfect_exp.return_value = True
+        result = await exp_cache_handler.get_one_perfect_exp()
+        assert result == "perfect_resp"
 
     @pytest.mark.asyncio
-    async def test_execute_function_when_no_perfect_exp(self, exp_cache_handler, mock_exp_manager, mock_func):
-        # Setup: No perfect experience
-        mock_exp_manager.extract_one_perfect_exp.return_value = None
-        mock_func.return_value = "Computed result"
-
-        # Exec
+    async def test_execute_function(self, exp_cache_handler, mock_func, mock_context_builder):
+        mock_context_builder.build.return_value = "built_context"
+        mock_func.return_value = "function_result"
         await exp_cache_handler.execute_function()
-
-        # Assert
-        assert exp_cache_handler._result == "Computed result"
-        mock_func.assert_called_once()
+        mock_context_builder.build.assert_called_once()
+        mock_func.assert_called_once_with(req="built_context")
+        assert exp_cache_handler._raw_resp == "function_result"
+        assert exp_cache_handler._resp == "function_result"
 
     @pytest.mark.asyncio
-    async def test_evaluate_and_save_experience(self, exp_cache_handler, mock_scorer, mock_exp_manager):
-        # Setup
-        mock_scorer.evaluate.return_value = Score(value=100)
-        exp_cache_handler._result = "Computed result"
-
-        # Exec
-        await exp_cache_handler.evaluate_experience()
-        exp_cache_handler.save_experience()
-
-        # Assert
+    async def test_process_experience(self, exp_cache_handler, mock_scorer, mock_exp_manager):
+        exp_cache_handler._resp = "test_resp"
+        mock_scorer.evaluate.return_value = Score(val=8)
+        await exp_cache_handler.process_experience()
         mock_scorer.evaluate.assert_called_once()
         mock_exp_manager.create_exp.assert_called_once()
 
     @pytest.mark.asyncio
-    async def test_async_function_execution_with_exps(self, exp_cache_handler, mock_exp_manager, mock_func):
-        # Setup
-        exp_cache_handler.pass_exps_to_func = True
-        mock_func.return_value = "Async result with exps"
-        mock_exp_manager.extract_one_perfect_exp.return_value = None
-        exp_cache_handler._exps = [Experience(req="req", resp="resp")]
+    async def test_evaluate_experience(self, exp_cache_handler, mock_scorer):
+        exp_cache_handler._resp = "test_resp"
+        mock_scorer.evaluate.return_value = Score(val=9)
+        await exp_cache_handler.evaluate_experience()
+        assert exp_cache_handler._score.val == 9
 
-        # Exec
-        await exp_cache_handler.execute_function()
-
-        # Assert
-        mock_func.assert_called_once_with(exps=exp_cache_handler._exps)
-        assert exp_cache_handler._result == "Async result with exps"
-
-    def test_sync_function_execution_with_exps(self, mocker, exp_cache_handler, mock_exp_manager, mock_func):
-        # Setup
-        exp_cache_handler.func = mocker.Mock(return_value="Sync result with exps")
-        exp_cache_handler.pass_exps_to_func = True
-        mock_exp_manager.extract_one_perfect_exp.return_value = None
-        exp_cache_handler._exps = [Experience(req="req", resp="resp")]
-
-        # Exec
-        asyncio.get_event_loop().run_until_complete(exp_cache_handler.execute_function())
-
-        # Assert
-        exp_cache_handler.func.assert_called_once_with(exps=exp_cache_handler._exps)
-        assert exp_cache_handler._result == "Sync result with exps"
-
-    def test_wrapper_selection_async(self, mocker, exp_cache_handler, mock_func):
-        # Setup
-        mock_func = mocker.AsyncMock()
-
-        # Exec
-        wrapper = ExpCacheHandler.choose_wrapper(mock_func, exp_cache_handler.execute_function)
-
-        # Assert
-        assert asyncio.iscoroutinefunction(wrapper), "Wrapper should be asynchronous"
-
-    def test_wrapper_selection_sync(self, exp_cache_handler, mocker):
-        # Setup
-        sync_func = mocker.Mock()
-
-        # Exec
-        wrapper = ExpCacheHandler.choose_wrapper(sync_func, exp_cache_handler.execute_function)
-
-        # Assert
-        assert not asyncio.iscoroutinefunction(wrapper), "Wrapper should be synchronous"
-
-    @pytest.mark.parametrize(
-        "func, args, kwargs, expected",
-        [
-            (for_test_function, (1, 2), {"c": 3}, 'for_test_function@[1~2]@{"c"!3}'),
-            (ForTestClass().for_test_method, (4, 5), {}, "ForTestClass.for_test_method@[4~5]@{}"),
-            (ForTestClass.for_test_class_method, (6, 7), {}, "ForTestClass.for_test_class_method@[6~7]@{}"),
-            (for_test_function, (), {}, "for_test_function@[]@{}"),
-            (
-                for_test_function,
-                ("hello", [1, 2]),
-                {"key": "value"},
-                'for_test_function@["hello"~[1~2]]@{"key"!"value"}',
-            ),
-        ],
-    )
-    def test_generate_req_identifier(self, func, args, kwargs, expected):
-        req_identifier = ExpCacheHandler.generate_req_identifier(func, *args, **kwargs)
-        assert req_identifier == expected
-
-    @pytest.mark.asyncio
-    async def test_exp_cache_with_perfect_experience(self, mocker, mock_exp_manager):
-        # Mock perfect experience
-        perfect_exp = Experience(req="test_req", resp="perfect_response")
-        mock_exp_manager.query_exps = mocker.AsyncMock(return_value=[perfect_exp])
-        mock_exp_manager.extract_one_perfect_exp = mocker.MagicMock(return_value=perfect_exp)
-        async_mock_func = mocker.AsyncMock()
-
-        # Setup
-        decorated_func = exp_cache(async_mock_func, manager=mock_exp_manager)
-
-        # Exec
-        result: Experience = await decorated_func()
-
-        # Assert
-        assert result.resp == "perfect_response", "Should return the perfect experience response"
-        async_mock_func.assert_not_called()
-
-    @pytest.mark.asyncio
-    async def test_exp_cache_without_perfect_experience(self, mocker, mock_exp_manager):
-        # Mock
-        mock_exp_manager.query_exps = mocker.AsyncMock(return_value=[])
-        mock_exp_manager.extract_one_perfect_exp = mocker.MagicMock(return_value=None)
-        async_mock_func = mocker.AsyncMock(return_value="computed_response")
-        async_mock_func.__signature__ = inspect.signature(for_test_function)
-
-        # Setup
-        decorated_func = exp_cache(async_mock_func, manager=mock_exp_manager)
-
-        # Exec
-        result = await decorated_func()
-
-        # Assert
-        assert result == "computed_response", "Should execute and return the function's response"
-        async_mock_func.assert_called_once()
-
-    @pytest.mark.asyncio
-    async def test_exp_cache_saves_new_experience(self, mocker, mock_exp_manager, mock_scorer):
-        # Mock
-        mock_exp_manager.query_exps = mocker.AsyncMock(return_value=[])
-        mock_exp_manager.extract_one_perfect_exp = mocker.MagicMock(return_value=None)
-        async_mock_func = mocker.AsyncMock(return_value="computed_response")
-        mock_scorer.evaluate = mocker.AsyncMock(return_value=Score(value=100))
-
-        # Setup
-        decorated_func = exp_cache(async_mock_func, manager=mock_exp_manager, scorer=mock_scorer)
-
-        # Exec
-        await decorated_func()
-
-        # Assert
+    def test_save_experience(self, exp_cache_handler, mock_exp_manager):
+        exp_cache_handler._req = "test_req"
+        exp_cache_handler._resp = "test_resp"
+        exp_cache_handler._score = Score(val=7)
+        exp_cache_handler.save_experience()
         mock_exp_manager.create_exp.assert_called_once()
+
+    def test_choose_wrapper_async(self, mocker):
+        async def async_func():
+            pass
+
+        wrapper = ExpCacheHandler.choose_wrapper(async_func, mocker.AsyncMock())
+        assert asyncio.iscoroutinefunction(wrapper)
+
+    def test_choose_wrapper_sync(self, mocker):
+        def sync_func():
+            pass
+
+        wrapper = ExpCacheHandler.choose_wrapper(sync_func, mocker.AsyncMock())
+        assert not asyncio.iscoroutinefunction(wrapper)
+
+    def test_validate_params(self):
+        with pytest.raises(ValueError):
+            ExpCacheHandler(func=lambda x: x, args=(), kwargs={})
+
+    def test_generate_tag(self):
+        class TestClass:
+            def test_method(self):
+                pass
+
+        handler = ExpCacheHandler(func=TestClass().test_method, args=(TestClass(),), kwargs={"req": "test"})
+        assert handler._generate_tag() == "TestClass.test_method"
+
+        handler = ExpCacheHandler(func=lambda x: x, args=(), kwargs={"req": "test"})
+        assert handler._generate_tag() == "<lambda>"
+
+
+class TestExpCache:
+    @pytest.fixture
+    def mock_exp_manager(self, mocker):
+        manager = mocker.MagicMock(spec=ExperienceManager)
+        manager.storage = mocker.MagicMock(spec=SimpleEngine)
+        manager.query_exps = mocker.AsyncMock()
+        manager.create_exp = mocker.MagicMock()
+        return manager
+
+    @pytest.fixture
+    def mock_scorer(self, mocker):
+        scorer = mocker.MagicMock(spec=SimpleScorer)
+        scorer.evaluate = mocker.AsyncMock(return_value=Score())
+        return scorer
+
+    @pytest.fixture
+    def mock_perfect_judge(self, mocker):
+        return mocker.MagicMock(spec=SimplePerfectJudge)
+
+    @pytest.fixture
+    def mock_config(self, mocker):
+        return mocker.patch("metagpt.exp_pool.decorator.config")
+
+    @pytest.mark.asyncio
+    async def test_exp_cache_disabled(self, mock_config, mock_exp_manager):
+        mock_config.exp_pool.enable_read = False
+
+        @exp_cache(manager=mock_exp_manager)
+        async def test_func(req):
+            return "result"
+
+        result = await test_func(req="test")
+        assert result == "result"
+        mock_exp_manager.query_exps.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_exp_cache_enabled_no_perfect_exp(self, mock_config, mock_exp_manager, mock_scorer):
+        mock_config.exp_pool.enable_read = True
+        mock_exp_manager.query_exps.return_value = []
+
+        @exp_cache(manager=mock_exp_manager, scorer=mock_scorer)
+        async def test_func(req):
+            return "computed_result"
+
+        result = await test_func(req="test")
+        assert result == "computed_result"
+        mock_exp_manager.query_exps.assert_called()
+        mock_exp_manager.create_exp.assert_called()
+
+    @pytest.mark.asyncio
+    async def test_exp_cache_enabled_with_perfect_exp(self, mock_config, mock_exp_manager, mock_perfect_judge):
+        mock_config.exp_pool.enable_read = True
+        perfect_exp = Experience(req="test", resp="perfect_result")
+        mock_exp_manager.query_exps.return_value = [perfect_exp]
+        mock_perfect_judge.is_perfect_exp.return_value = True
+
+        @exp_cache(manager=mock_exp_manager, perfect_judge=mock_perfect_judge)
+        async def test_func(req):
+            return "should_not_be_called"
+
+        result = await test_func(req="test")
+        assert result == "perfect_result"
+        mock_exp_manager.query_exps.assert_called_once()
+        mock_exp_manager.create_exp.assert_not_called()
diff --git a/tests/metagpt/exp_pool/test_manager.py b/tests/metagpt/exp_pool/test_manager.py
index 3e8f47417..c12fc7e8c 100644
--- a/tests/metagpt/exp_pool/test_manager.py
+++ b/tests/metagpt/exp_pool/test_manager.py
@@ -4,20 +4,25 @@ from metagpt.config2 import Config
 from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 from metagpt.configs.llm_config import LLMConfig
 from metagpt.exp_pool.manager import ExperienceManager
-from metagpt.exp_pool.schema import MAX_SCORE, Experience, Metric, Score
+from metagpt.exp_pool.schema import Experience
 from metagpt.rag.engines import SimpleEngine
 
 
 class TestExperienceManager:
     @pytest.fixture
     def mock_config(self):
-        return Config(llm=LLMConfig(), exp_pool=ExperiencePoolConfig(enable_write=True, enable_read=True))
+        return Config(
+            llm=LLMConfig(), exp_pool=ExperiencePoolConfig(enable_write=True, enable_read=True, init_exp=False)
+        )
 
     @pytest.fixture
     def mock_storage(self, mocker):
         engine = mocker.MagicMock(spec=SimpleEngine)
         engine.add_objs = mocker.MagicMock()
         engine.aretrieve = mocker.AsyncMock(return_value=[])
+        engine._retriever = mocker.MagicMock()
+        engine._retriever._vector_store = mocker.MagicMock()
+        engine._retriever._vector_store._get = mocker.MagicMock(return_value=mocker.MagicMock(ids=[]))
         return engine
 
     @pytest.fixture
@@ -33,7 +38,7 @@ class TestExperienceManager:
 
     def test_create_exp(self, mock_experience_manager, mock_experience):
         mock_experience_manager.create_exp(mock_experience)
-        mock_experience_manager.storage.add_objs.assert_called_once_with([mock_experience])
+        mock_experience_manager.storage.add_objs.assert_called_with([mock_experience])
 
     def test_create_exp_write_disabled(self, mock_experience_manager, mock_experience, mock_config):
         mock_config.exp_pool.enable_write = False
@@ -60,18 +65,44 @@ class TestExperienceManager:
         result = await mock_experience_manager.query_exps("query")
         assert result == []
 
-    def test_extract_one_perfect_exp(self, mock_experience_manager):
-        experiences = [
-            Experience(req="req", resp="resp", metric=Metric(score=Score(val=MAX_SCORE))),
-            Experience(req="req", resp="resp"),
-        ]
-        perfect_exp: Experience = mock_experience_manager.extract_one_perfect_exp(experiences)
-        assert perfect_exp is not None
-        assert perfect_exp.metric.score.val == MAX_SCORE
+    def test_init_exp_pool(self, mock_experience_manager, mock_config, mocker):
+        mock_experience_manager._has_exps = mocker.MagicMock(return_value=False)
+        mock_experience_manager._init_teamleader_exps = mocker.MagicMock()
+        mock_experience_manager._init_engineer2_exps = mocker.MagicMock()
 
-    def test_is_perfect_exp(self):
-        exp = Experience(req="req", resp="resp", metric=Metric(score=Score(val=MAX_SCORE)))
-        assert ExperienceManager.is_perfect_exp(exp) == True
+        mock_config.exp_pool.init_exp = True
+        mock_experience_manager.init_exp_pool()
 
-        exp = Experience(req="req", resp="resp")
-        assert ExperienceManager.is_perfect_exp(exp) == False
+        mock_experience_manager._has_exps.assert_called_once()
+        mock_experience_manager._init_teamleader_exps.assert_called_once()
+        mock_experience_manager._init_engineer2_exps.assert_called_once()
+
+    def test_init_exp_pool_already_has_exps(self, mock_experience_manager, mock_config, mocker):
+        mock_experience_manager._has_exps = mocker.MagicMock(return_value=True)
+        mock_experience_manager._init_teamleader_exps = mocker.MagicMock()
+        mock_experience_manager._init_engineer2_exps = mocker.MagicMock()
+
+        mock_config.exp_pool.init_exp = True
+        mock_experience_manager.init_exp_pool()
+
+        mock_experience_manager._has_exps.assert_called_once()
+        mock_experience_manager._init_teamleader_exps.assert_not_called()
+        mock_experience_manager._init_engineer2_exps.assert_not_called()
+
+    def test_has_exps(self, mock_experience_manager, mock_storage):
+        mock_storage._retriever._vector_store._get.return_value.ids = ["id1"]
+
+        assert mock_experience_manager._has_exps() is True
+
+        mock_storage._retriever._vector_store._get.return_value.ids = []
+        assert mock_experience_manager._has_exps() is False
+
+    def test_init_teamleader_exps(self, mock_experience_manager, mocker):
+        mock_experience_manager._init_exp = mocker.MagicMock()
+        mock_experience_manager._init_teamleader_exps()
+        mock_experience_manager._init_exp.assert_called_once()
+
+    def test_init_engineer2_exps(self, mock_experience_manager, mocker):
+        mock_experience_manager._init_exp = mocker.MagicMock()
+        mock_experience_manager._init_engineer2_exps()
+        mock_experience_manager._init_exp.assert_called_once()
diff --git a/tests/metagpt/exp_pool/test_perfect_judges/test_simple_perfect_judge.py b/tests/metagpt/exp_pool/test_perfect_judges/test_simple_perfect_judge.py
new file mode 100644
index 000000000..5abd04f0d
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_perfect_judges/test_simple_perfect_judge.py
@@ -0,0 +1,40 @@
+import pytest
+
+from metagpt.exp_pool.perfect_judges import SimplePerfectJudge
+from metagpt.exp_pool.schema import MAX_SCORE, Experience, Metric, Score
+
+
+class TestSimplePerfectJudge:
+    @pytest.fixture
+    def simple_perfect_judge(self):
+        return SimplePerfectJudge()
+
+    @pytest.mark.asyncio
+    async def test_is_perfect_exp_perfect_match(self, simple_perfect_judge):
+        exp = Experience(req="test_request", resp="resp", metric=Metric(score=Score(val=MAX_SCORE)))
+        result = await simple_perfect_judge.is_perfect_exp(exp, "test_request")
+        assert result is True
+
+    @pytest.mark.asyncio
+    async def test_is_perfect_exp_imperfect_score(self, simple_perfect_judge):
+        exp = Experience(req="test_request", resp="resp", metric=Metric(score=Score(val=MAX_SCORE - 1)))
+        result = await simple_perfect_judge.is_perfect_exp(exp, "test_request")
+        assert result is False
+
+    @pytest.mark.asyncio
+    async def test_is_perfect_exp_mismatched_request(self, simple_perfect_judge):
+        exp = Experience(req="test_request", resp="resp", metric=Metric(score=Score(val=MAX_SCORE)))
+        result = await simple_perfect_judge.is_perfect_exp(exp, "different_request")
+        assert result is False
+
+    @pytest.mark.asyncio
+    async def test_is_perfect_exp_no_metric(self, simple_perfect_judge):
+        exp = Experience(req="test_request", resp="resp")
+        result = await simple_perfect_judge.is_perfect_exp(exp, "test_request")
+        assert result is False
+
+    @pytest.mark.asyncio
+    async def test_is_perfect_exp_no_score(self, simple_perfect_judge):
+        exp = Experience(req="test_request", resp="resp", metric=Metric())
+        result = await simple_perfect_judge.is_perfect_exp(exp, "test_request")
+        assert result is False
diff --git a/tests/metagpt/exp_pool/test_scorers/test_simple_scorer.py b/tests/metagpt/exp_pool/test_scorers/test_simple_scorer.py
new file mode 100644
index 000000000..043f105d0
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_scorers/test_simple_scorer.py
@@ -0,0 +1,49 @@
+import pytest
+
+from metagpt.exp_pool.schema import Score
+from metagpt.exp_pool.scorers.simple import SIMPLE_SCORER_TEMPLATE, SimpleScorer
+from metagpt.llm import BaseLLM
+
+
+class TestSimpleScorer:
+    @pytest.fixture
+    def mock_llm(self, mocker):
+        mock_llm = mocker.MagicMock(spec=BaseLLM)
+        return mock_llm
+
+    @pytest.fixture
+    def simple_scorer(self, mock_llm):
+        return SimpleScorer(llm=mock_llm)
+
+    def test_init(self, mock_llm):
+        scorer = SimpleScorer(llm=mock_llm)
+        assert isinstance(scorer.llm, BaseLLM)
+
+    @pytest.mark.asyncio
+    async def test_evaluate(self, simple_scorer, mock_llm):
+        # Mock function to evaluate
+        def mock_func(a, b):
+            """This is a mock function."""
+            return a + b
+
+        # Mock LLM response
+        mock_llm.aask.return_value = '```json\n{"val": 8, "reason": "Good performance"}\n```'
+
+        # Test evaluate method
+        result = await simple_scorer.evaluate(mock_func, 5, args=(2, 3), kwargs={})
+
+        # Assert LLM was called with correct prompt
+        expected_prompt = SIMPLE_SCORER_TEMPLATE.format(
+            func_name=mock_func.__name__,
+            func_doc=mock_func.__doc__,
+            func_signature="(a, b)",
+            func_args=(2, 3),
+            func_kwargs={},
+            func_result=5,
+        )
+        mock_llm.aask.assert_called_once_with(expected_prompt)
+
+        # Assert the result is correct
+        assert isinstance(result, Score)
+        assert result.val == 8
+        assert result.reason == "Good performance"

From 3fb1432158e6240b4546580de608d684e4109767 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 8 Jul 2024 10:50:20 +0800
Subject: [PATCH 32/51] delete unnecessary file

---
 tests/metagpt/utils/test_reflection.py | 35 --------------------------
 1 file changed, 35 deletions(-)
 delete mode 100644 tests/metagpt/utils/test_reflection.py

diff --git a/tests/metagpt/utils/test_reflection.py b/tests/metagpt/utils/test_reflection.py
deleted file mode 100644
index 58fd81619..000000000
--- a/tests/metagpt/utils/test_reflection.py
+++ /dev/null
@@ -1,35 +0,0 @@
-from metagpt.utils.reflection import get_class_name
-
-
-class SimpleFunction:
-    def function(self):
-        pass
-
-
-class SampleClass:
-    @classmethod
-    def class_method(cls):
-        pass
-
-    def instance_method(self):
-        pass
-
-
-def standalone_function():
-    pass
-
-
-class TestGetClassName:
-    def test_instance_method(self):
-        instance = SampleClass()
-        assert get_class_name(instance.instance_method) == "SampleClass"
-
-    def test_class_method(self):
-        assert get_class_name(SampleClass.class_method) == "SampleClass"
-
-    def test_standalone_function(self):
-        assert get_class_name(standalone_function) == ""
-
-    def test_function_within_simple_class(self):
-        instance = SimpleFunction()
-        assert get_class_name(instance.function) == "SimpleFunction"

From 361294d31d4f2d478e7f9149cf2be08670afb261 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 8 Jul 2024 11:11:07 +0800
Subject: [PATCH 33/51] delete unnecessary code

---
 metagpt/exp_pool/schema.py | 16 ----------------
 1 file changed, 16 deletions(-)

diff --git a/metagpt/exp_pool/schema.py b/metagpt/exp_pool/schema.py
index d59478742..627dcbb4e 100644
--- a/metagpt/exp_pool/schema.py
+++ b/metagpt/exp_pool/schema.py
@@ -3,7 +3,6 @@
 from enum import Enum
 from typing import Optional
 
-from llama_index.core.schema import TextNode
 from pydantic import BaseModel, Field
 
 MAX_SCORE = 10
@@ -71,18 +70,3 @@ class Experience(BaseModel):
 
     def rag_key(self):
         return self.req
-
-
-class ExperienceNodeMetadata(BaseModel):
-    """Metadata of ExperienceNode."""
-
-    resp: str = Field(..., description="")
-
-
-class ExperienceNode(TextNode):
-    """ExperienceNode for RAG."""
-
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self.excluded_llm_metadata_keys = list(ExperienceNodeMetadata.model_fields.keys())
-        self.excluded_embed_metadata_keys = self.excluded_llm_metadata_keys

From a2bb67a1f0cf829e3856a1a128edb03fd25377d5 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 8 Jul 2024 15:24:18 +0800
Subject: [PATCH 34/51] move the req_serialize from di/role_zero.py to
 context_builders/role_zero.py

---
 .../exp_pool/context_builders/role_zero.py    | 24 +++++++++++++++++
 metagpt/exp_pool/decorator.py                 |  1 -
 metagpt/exp_pool/manager.py                   |  1 +
 metagpt/roles/di/role_zero.py                 | 27 +++----------------
 4 files changed, 28 insertions(+), 25 deletions(-)

diff --git a/metagpt/exp_pool/context_builders/role_zero.py b/metagpt/exp_pool/context_builders/role_zero.py
index 60f71ef59..e9ab83d90 100644
--- a/metagpt/exp_pool/context_builders/role_zero.py
+++ b/metagpt/exp_pool/context_builders/role_zero.py
@@ -1,4 +1,6 @@
 """RoleZero context builder."""
+import copy
+import json
 
 from metagpt.exp_pool.context_builders.base import BaseContextBuilder
 
@@ -24,3 +26,25 @@ class RoleZeroContextBuilder(BaseContextBuilder):
 
     def replace_example_content(self, text: str, new_example_content: str) -> str:
         return self.replace_content_between_markers(text, "# Example", "# Instruction", new_example_content)
+
+    @staticmethod
+    def req_serialize(req: list[dict]) -> str:
+        """Serialize the request for database storage, ensuring it is a string.
+
+        This function deep copies the request and modifies the content of the last element
+        to remove unnecessary sections, making the request more concise.
+        """
+
+        req_copy = copy.deepcopy(req)
+
+        last_content = req_copy[-1]["content"]
+        last_content = RoleZeroContextBuilder.replace_content_between_markers(
+            last_content, "# Data Structure", "# Current Plan", ""
+        )
+        last_content = RoleZeroContextBuilder.replace_content_between_markers(
+            last_content, "# Example", "# Instruction", ""
+        )
+
+        req_copy[-1]["content"] = last_content
+
+        return json.dumps(req_copy)
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index c518bb7ea..10f3355f9 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -56,7 +56,6 @@ def exp_cache(
 
         @functools.wraps(func)
         async def get_or_create(args: Any, kwargs: Any) -> ReturnType:
-            logger.info("exp_cache is enabled.")
             handler = ExpCacheHandler(
                 func=func,
                 args=args,
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 276b1e8e3..23198eb02 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -49,6 +49,7 @@ class ExperienceManager(BaseModel):
 
         self.init_exp_pool()
 
+        logger.debug(f"exp_pool config: {self.config.exp_pool}")
         return self
 
     @handle_exception
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 2c2c81551..103a77911 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -1,6 +1,5 @@
 from __future__ import annotations
 
-import copy
 import inspect
 import json
 import re
@@ -166,32 +165,12 @@ class RoleZero(Role):
 
         return True
 
-    @exp_cache(context_builder=RoleZeroContextBuilder(), req_serialize=lambda req: RoleZero._req_serialize(req))
+    @exp_cache(
+        context_builder=RoleZeroContextBuilder(), req_serialize=lambda req: RoleZeroContextBuilder.req_serialize(req)
+    )
     async def llm_cached_aask(self, *, req: list[dict], system_msgs: list[str]) -> str:
         return await self.llm.aask(req, system_msgs=system_msgs)
 
-    @staticmethod
-    def _req_serialize(req: list[dict]) -> str:
-        """Serialize the request for database storage, ensuring it is a string.
-
-        This function deep copies the request and modifies the content of the last element
-        to remove unnecessary sections, making the request more concise.
-        """
-
-        req_copy = copy.deepcopy(req)
-
-        last_content = req_copy[-1]["content"]
-        last_content = RoleZeroContextBuilder.replace_content_between_markers(
-            last_content, "# Data Structure", "# Current Plan", ""
-        )
-        last_content = RoleZeroContextBuilder.replace_content_between_markers(
-            last_content, "# Example", "# Instruction", ""
-        )
-
-        req_copy[-1]["content"] = last_content
-
-        return json.dumps(req_copy)
-
     async def _act(self) -> Message:
         if self.use_fixed_sop:
             return await super()._act()

From f61506bd3220252351882cbdb8e0e28cf5513b13 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 8 Jul 2024 16:24:37 +0800
Subject: [PATCH 35/51] update format_exps

---
 metagpt/exp_pool/context_builders/base.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/metagpt/exp_pool/context_builders/base.py b/metagpt/exp_pool/context_builders/base.py
index e3fe320a6..a261e452e 100644
--- a/metagpt/exp_pool/context_builders/base.py
+++ b/metagpt/exp_pool/context_builders/base.py
@@ -25,7 +25,8 @@ class BaseContextBuilder(BaseModel, ABC):
 
         result = []
         for i, exp in enumerate(self.exps, start=1):
-            result.append(f"{i}. " + EXP_TEMPLATE.format(req=exp.req, resp=exp.resp, score=exp.metric.score.val))
+            score_val = exp.metric.score.val if exp.metric and exp.metric.score else "N/A"
+            result.append(f"{i}. " + EXP_TEMPLATE.format(req=exp.req, resp=exp.resp, score=score_val))
 
         return "\n".join(result)
 

From 1ead3e4d8083c258d0d418eb3cfab3564504a188 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 8 Jul 2024 20:55:32 +0800
Subject: [PATCH 36/51] update simple_scorer

---
 examples/exp_pool/scorer.py                   | 19 ++++---
 metagpt/exp_pool/decorator.py                 |  2 +-
 metagpt/exp_pool/scorers/base.py              | 16 +-----
 metagpt/exp_pool/scorers/simple.py            | 53 ++++++-------------
 tests/metagpt/exp_pool/test_decorator.py      |  7 ++-
 .../test_scorers/test_simple_scorer.py        | 49 +++++++++++------
 6 files changed, 71 insertions(+), 75 deletions(-)

diff --git a/examples/exp_pool/scorer.py b/examples/exp_pool/scorer.py
index 1efe07bdf..c412feaf3 100644
--- a/examples/exp_pool/scorer.py
+++ b/examples/exp_pool/scorer.py
@@ -1,20 +1,27 @@
 import asyncio
 
 from metagpt.exp_pool.scorers import SimpleScorer
-from metagpt.logs import logger
 
+REQ = "Write a program to implement quicksort in python."
 
-def echo(req: str):
-    """Echo from req."""
+RESP1 = """
+def quicksort(arr):
+    return quicksort([x for x in arr[1:] if x <= arr[0]]) + [arr[0]] + quicksort([x for x in arr[1:] if x > arr[0]])
+"""
 
-    return req
+RESP2 = """
+def quicksort(arr):
+    if len(arr) <= 1:
+        return arr
+    return quicksort([x for x in arr[1:] if x <= arr[0]]) + [arr[0]] + quicksort([x for x in arr[1:] if x > arr[0]])
+"""
 
 
 async def simple():
     scorer = SimpleScorer()
 
-    score = await scorer.evaluate(echo, "data", ("data",))
-    logger.info(f"The score is: {score}")
+    await scorer.evaluate(req=REQ, resp=RESP1)
+    await scorer.evaluate(req=REQ, resp=RESP2)
 
 
 async def main():
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 10f3355f9..4e7213dfe 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -159,7 +159,7 @@ class ExpCacheHandler(BaseModel):
     async def evaluate_experience(self):
         """Evaluate the experience, and save the score."""
 
-        self._score = await self.exp_scorer.evaluate(self.func, self._resp, self.args, self.kwargs)
+        self._score = await self.exp_scorer.evaluate(self._req, self._resp)
 
     def save_experience(self):
         """Save the new experience."""
diff --git a/metagpt/exp_pool/scorers/base.py b/metagpt/exp_pool/scorers/base.py
index 94623c30f..97cac4992 100644
--- a/metagpt/exp_pool/scorers/base.py
+++ b/metagpt/exp_pool/scorers/base.py
@@ -1,7 +1,6 @@
 """Base scorer."""
 
 from abc import ABC, abstractmethod
-from typing import Any, Callable
 
 from pydantic import BaseModel, ConfigDict
 
@@ -12,16 +11,5 @@ class BaseScorer(BaseModel, ABC):
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
     @abstractmethod
-    async def evaluate(self, func: Callable, result: Any, args: tuple = None, kwargs: dict = None) -> Score:
-        """Evaluate the quality of the result produced by the function and parameters.
-
-        Args:
-            func (Callable): The function whose result is to be evaluated.
-            result (Any): The result produced by the function.
-            args (Tuple[Any, ...]): The tuple of arguments that were passed to the function.
-            kwargs (Dict[str, Any]): The dictionary of keyword arguments that were passed to the function.
-
-        Example:
-            result = await sample(5, name="foo")
-            score = await scorer.evaluate(sample, result, args=(5), kwargs={"name": "foo"})
-        """
+    async def evaluate(self, req: str, resp: str) -> Score:
+        """Evaluates the quality of a response relative to a given request."""
diff --git a/metagpt/exp_pool/scorers/simple.py b/metagpt/exp_pool/scorers/simple.py
index 1fda189d1..fd7b6537b 100644
--- a/metagpt/exp_pool/scorers/simple.py
+++ b/metagpt/exp_pool/scorers/simple.py
@@ -1,8 +1,6 @@
 """Simple scorer."""
 
-import inspect
 import json
-from typing import Any, Callable
 
 from pydantic import Field
 
@@ -13,24 +11,16 @@ from metagpt.provider.base_llm import BaseLLM
 from metagpt.utils.common import CodeParser
 
 SIMPLE_SCORER_TEMPLATE = """
-Role: You're an expert score evaluator. You specialize in assessing the output of the given function, based on its intended requirement and produced result.
+Role: You are a highly efficient assistant, tasked with evaluating a response to a given request. The response is generated by a large language model (LLM). 
+
+I will provide you with a request and a corresponding response. Your task is to assess this response and provide a score from a human perspective.
 
 ## Context
-### Function Name
-{func_name}
+### Request
+{req}
 
-### Function Document
-{func_doc}
-
-### Function Signature
-{func_signature}
-
-### Function Parameters
-args: {func_args}
-kwargs: {func_kwargs}
-
-### Produced Result By Function and Parameters
-{func_result}
+### Response
+{resp}
 
 ## Format Example
 ```json
@@ -41,10 +31,10 @@ kwargs: {func_kwargs}
 ```
 
 ## Instructions
-- Understand the function and requirements given by the user.
-- Analyze the results produced by the function.
-- Grade the results based on level of alignment with the requirements.
-- Provide a score on a scale defined by user or a default scale (1 to 10).
+- Understand the request and response given by the user.
+- Evaluate the response based on its quality relative to the given request.
+- Provide a score from 1 to 10, where 10 is the best.
+- Provide a reason supporting your score.
 
 ## Constraint
 Format: Just print the result in json format like **Format Example**.
@@ -57,26 +47,17 @@ Follow instructions, generate output and make sure it follows the **Constraint**
 class SimpleScorer(BaseScorer):
     llm: BaseLLM = Field(default_factory=LLM)
 
-    async def evaluate(self, func: Callable, result: Any, args: tuple = None, kwargs: dict = None) -> Score:
-        """Evaluates the quality of content by LLM.
+    async def evaluate(self, req: str, resp: str) -> Score:
+        """Evaluates the quality of a response relative to a given request, as scored by an LLM.
 
         Args:
-            func: The function to evaluate.
-            result: The result produced by the function.
-            args: The positional arguments used when calling the function, if any.
-            kwargs: The keyword arguments used when calling the function, if any.
+            req (str): The request.
+            resp (str): The response.
 
         Returns:
-            A Score object containing the evaluation results.
+            Score: An object containing the score (1-10) and the reasoning.
         """
-        prompt = SIMPLE_SCORER_TEMPLATE.format(
-            func_name=func.__name__,
-            func_doc=func.__doc__,
-            func_signature=inspect.signature(func),
-            func_args=args,
-            func_kwargs=kwargs,
-            func_result=result,
-        )
+        prompt = SIMPLE_SCORER_TEMPLATE.format(req=req, resp=resp)
         resp = await self.llm.aask(prompt)
         resp_json = json.loads(CodeParser.parse_code(resp, lang="json"))
 
diff --git a/tests/metagpt/exp_pool/test_decorator.py b/tests/metagpt/exp_pool/test_decorator.py
index c0b3fe36d..0c02dcdfc 100644
--- a/tests/metagpt/exp_pool/test_decorator.py
+++ b/tests/metagpt/exp_pool/test_decorator.py
@@ -2,6 +2,8 @@ import asyncio
 
 import pytest
 
+from metagpt.config2 import Config
+from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 from metagpt.exp_pool.context_builders import SimpleContextBuilder
 from metagpt.exp_pool.decorator import ExpCacheHandler, exp_cache
 from metagpt.exp_pool.manager import ExperienceManager
@@ -20,6 +22,8 @@ class TestExpCacheHandler:
     def mock_exp_manager(self, mocker):
         manager = mocker.MagicMock(spec=ExperienceManager)
         manager.storage = mocker.MagicMock(spec=SimpleEngine)
+        manager.config = mocker.MagicMock(spec=Config)
+        manager.config.exp_pool = ExperiencePoolConfig()
         manager.query_exps = mocker.AsyncMock()
         manager.create_exp = mocker.MagicMock()
         return manager
@@ -131,9 +135,10 @@ class TestExpCacheHandler:
 
 class TestExpCache:
     @pytest.fixture
-    def mock_exp_manager(self, mocker):
+    def mock_exp_manager(self, mocker, mock_config):
         manager = mocker.MagicMock(spec=ExperienceManager)
         manager.storage = mocker.MagicMock(spec=SimpleEngine)
+        manager.config = mock_config
         manager.query_exps = mocker.AsyncMock()
         manager.create_exp = mocker.MagicMock()
         return manager
diff --git a/tests/metagpt/exp_pool/test_scorers/test_simple_scorer.py b/tests/metagpt/exp_pool/test_scorers/test_simple_scorer.py
index 043f105d0..e17edfca8 100644
--- a/tests/metagpt/exp_pool/test_scorers/test_simple_scorer.py
+++ b/tests/metagpt/exp_pool/test_scorers/test_simple_scorer.py
@@ -1,3 +1,5 @@
+import json
+
 import pytest
 
 from metagpt.exp_pool.schema import Score
@@ -20,30 +22,43 @@ class TestSimpleScorer:
         assert isinstance(scorer.llm, BaseLLM)
 
     @pytest.mark.asyncio
-    async def test_evaluate(self, simple_scorer, mock_llm):
-        # Mock function to evaluate
-        def mock_func(a, b):
-            """This is a mock function."""
-            return a + b
+    async def test_evaluate(self, simple_scorer, mock_llm, mocker):
+        # Mock request and response
+        req = "What is the capital of France?"
+        resp = "The capital of France is Paris."
 
         # Mock LLM response
-        mock_llm.aask.return_value = '```json\n{"val": 8, "reason": "Good performance"}\n```'
+        mock_llm_response = '{"val": 9, "reason": "Accurate and concise answer"}'
+        mock_llm.aask.return_value = f"```json\n{mock_llm_response}\n```"
+
+        # Mock CodeParser.parse_code
+        mocker.patch("metagpt.utils.common.CodeParser.parse_code", return_value=mock_llm_response)
 
         # Test evaluate method
-        result = await simple_scorer.evaluate(mock_func, 5, args=(2, 3), kwargs={})
+        result = await simple_scorer.evaluate(req, resp)
 
         # Assert LLM was called with correct prompt
-        expected_prompt = SIMPLE_SCORER_TEMPLATE.format(
-            func_name=mock_func.__name__,
-            func_doc=mock_func.__doc__,
-            func_signature="(a, b)",
-            func_args=(2, 3),
-            func_kwargs={},
-            func_result=5,
-        )
+        expected_prompt = SIMPLE_SCORER_TEMPLATE.format(req=req, resp=resp)
         mock_llm.aask.assert_called_once_with(expected_prompt)
 
         # Assert the result is correct
         assert isinstance(result, Score)
-        assert result.val == 8
-        assert result.reason == "Good performance"
+        assert result.val == 9
+        assert result.reason == "Accurate and concise answer"
+
+    @pytest.mark.asyncio
+    async def test_evaluate_invalid_response(self, simple_scorer, mock_llm, mocker):
+        # Mock request and response
+        req = "What is the capital of France?"
+        resp = "The capital of France is Paris."
+
+        # Mock LLM response with invalid JSON
+        mock_llm_response = "Invalid JSON"
+        mock_llm.aask.return_value = f"```json\n{mock_llm_response}\n```"
+
+        # Mock CodeParser.parse_code
+        mocker.patch("metagpt.utils.common.CodeParser.parse_code", return_value=mock_llm_response)
+
+        # Test evaluate method with invalid response
+        with pytest.raises(json.JSONDecodeError):
+            await simple_scorer.evaluate(req, resp)

From 086ef5e8055e0ad43833bfcd985bb46a801cf1d7 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 9 Jul 2024 12:41:44 +0800
Subject: [PATCH 37/51] update comment

---
 metagpt/exp_pool/decorator.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 4e7213dfe..62f766b9d 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -109,6 +109,11 @@ class ExpCacheHandler(BaseModel):
 
     @model_validator(mode="after")
     def initialize(self):
+        """Initialize default values for optional parameters if they are None.
+
+        This is necessary because the decorator might pass None, which would override the default values set by Field.
+        """
+
         self._validate_params()
 
         self.exp_manager = self.exp_manager or exp_manager

From b5934a412bbaaee25337b60ac6c4610d5077c324 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Wed, 10 Jul 2024 10:24:04 +0800
Subject: [PATCH 38/51] add serializers to support serialization and
 deserialization.

---
 config/config2.example.yaml                   |  1 -
 metagpt/actions/action_node.py                | 25 +----------
 metagpt/configs/exp_pool_config.py            |  3 --
 .../exp_pool/context_builders/action_node.py  | 33 +++++++++++++++
 metagpt/exp_pool/context_builders/base.py     | 35 +++++-----------
 .../exp_pool/context_builders/role_zero.py    | 42 ++++++++++---------
 metagpt/exp_pool/context_builders/simple.py   | 12 +++---
 metagpt/exp_pool/decorator.py                 | 30 +++++--------
 metagpt/exp_pool/manager.py                   |  4 +-
 metagpt/exp_pool/serializers/__init__.py      |  9 ++++
 metagpt/exp_pool/serializers/action_node.py   | 36 ++++++++++++++++
 metagpt/exp_pool/serializers/base.py          | 22 ++++++++++
 metagpt/exp_pool/serializers/role_zero.py     | 40 ++++++++++++++++++
 metagpt/exp_pool/serializers/simple.py        | 22 ++++++++++
 metagpt/roles/di/role_zero.py                 |  5 +--
 .../test_base_context_builder.py              | 13 ------
 .../test_rolezero_context_builder.py          | 15 ++++++-
 .../test_simple_context_builder.py            |  3 +-
 tests/metagpt/exp_pool/test_manager.py        | 28 +------------
 19 files changed, 234 insertions(+), 144 deletions(-)
 create mode 100644 metagpt/exp_pool/context_builders/action_node.py
 create mode 100644 metagpt/exp_pool/serializers/__init__.py
 create mode 100644 metagpt/exp_pool/serializers/action_node.py
 create mode 100644 metagpt/exp_pool/serializers/base.py
 create mode 100644 metagpt/exp_pool/serializers/role_zero.py
 create mode 100644 metagpt/exp_pool/serializers/simple.py

diff --git a/config/config2.example.yaml b/config/config2.example.yaml
index a3bd5c367..330b73680 100644
--- a/config/config2.example.yaml
+++ b/config/config2.example.yaml
@@ -78,7 +78,6 @@ exp_pool:
   enable_read: false 
   enable_write: false
   persist_path: .chroma_exp_data # The directory.
-  init_exp: false # If set to true, basic experiences associated with the roles will be added to the experience pool.
 
 azure_tts_subscription_key: "YOUR_SUBSCRIPTION_KEY"
 azure_tts_region: "eastus"
diff --git a/metagpt/actions/action_node.py b/metagpt/actions/action_node.py
index e1e0bddbb..c1de16656 100644
--- a/metagpt/actions/action_node.py
+++ b/metagpt/actions/action_node.py
@@ -19,6 +19,7 @@ from tenacity import retry, stop_after_attempt, wait_random_exponential
 from metagpt.actions.action_outcls_registry import register_action_outcls
 from metagpt.const import MARKDOWN_TITLE_PREFIX, USE_CONFIG_TIMEOUT
 from metagpt.exp_pool import exp_cache
+from metagpt.exp_pool.serializers import ActionNodeSerializer
 from metagpt.llm import BaseLLM
 from metagpt.logs import logger
 from metagpt.provider.postprocess.llm_output_postprocess import llm_output_postprocess
@@ -466,29 +467,7 @@ class ActionNode:
 
         return self
 
-    @classmethod
-    def deserialize_to_action_node(cls, serialized_data) -> "ActionNode":
-        """Customized deserialization, it will be triggered when a perfect experience is found.
-
-        ActionNode cannot be serialized, it throws an error 'cannot pickle 'SSLContext' object'.
-        """
-
-        class InstructContent:
-            def __init__(self, json_data):
-                self.json_data = json_data
-
-            def model_dump_json(self):
-                return self.json_data
-
-        action_node = cls(key="", expected_type=Type[str], instruction="", example="")
-        action_node.instruct_content = InstructContent(serialized_data)
-
-        return action_node
-
-    @exp_cache(
-        resp_serialize=lambda action_node: action_node.instruct_content.model_dump_json(),
-        resp_deserialize=lambda resp: ActionNode.deserialize_to_action_node(resp),
-    )
+    @exp_cache(serializer=ActionNodeSerializer())
     async def fill(
         self,
         *,
diff --git a/metagpt/configs/exp_pool_config.py b/metagpt/configs/exp_pool_config.py
index 0c92312da..786558ed9 100644
--- a/metagpt/configs/exp_pool_config.py
+++ b/metagpt/configs/exp_pool_config.py
@@ -7,6 +7,3 @@ class ExperiencePoolConfig(YamlModel):
     enable_read: bool = Field(default=False, description="Enable to read from experience pool.")
     enable_write: bool = Field(default=False, description="Enable to write to experience pool.")
     persist_path: str = Field(default=".chroma_exp_data", description="The persist path for experience pool.")
-    init_exp: bool = Field(
-        default=False, description="Put some basic experiences associated with the roles into the experience pool."
-    )
diff --git a/metagpt/exp_pool/context_builders/action_node.py b/metagpt/exp_pool/context_builders/action_node.py
new file mode 100644
index 000000000..ade157822
--- /dev/null
+++ b/metagpt/exp_pool/context_builders/action_node.py
@@ -0,0 +1,33 @@
+"""Action Node context builder."""
+
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+
+ACTION_NODE_CONTEXT_TEMPLATE = """
+{req}
+
+### Experiences
+-----
+{exps}
+-----
+
+## Instruction
+Consider **Experiences** to generate a better answer.
+"""
+
+
+class ActionNodeContextBuilder(BaseContextBuilder):
+    async def build(self, **kwargs) -> str:
+        """Builds the action node context string.
+
+        Args:
+            **kwargs: Arbitrary keyword arguments, expecting 'req' as a key.
+
+        Returns:
+            str: The formatted context string using the request and formatted experiences.
+                 If no experiences are available, returns the request as is.
+        """
+        req = kwargs.get("req", "")
+        exps = self.format_exps()
+
+        return ACTION_NODE_CONTEXT_TEMPLATE.format(req=req, exps=exps) if exps else req
diff --git a/metagpt/exp_pool/context_builders/base.py b/metagpt/exp_pool/context_builders/base.py
index a261e452e..d1133c2da 100644
--- a/metagpt/exp_pool/context_builders/base.py
+++ b/metagpt/exp_pool/context_builders/base.py
@@ -1,6 +1,5 @@
 """Base context builder."""
 
-import re
 from abc import ABC, abstractmethod
 from typing import Any
 
@@ -17,11 +16,19 @@ class BaseContextBuilder(BaseModel, ABC):
     exps: list[Experience] = []
 
     @abstractmethod
-    async def build(self, *args, **kwargs) -> Any:
+    async def build(self, **kwargs) -> Any:
         """Build context from parameters."""
 
     def format_exps(self) -> str:
-        """Format experiences into a numbered list of strings."""
+        """Format experiences into a numbered list of strings.
+
+        Example:
+            1. Given the request: req1, We can get the response: resp1, Which scored: 8.
+            2. Given the request: req2, We can get the response: resp2, Which scored: 9.
+
+        Returns:
+            str: The formatted experiences as a string.
+        """
 
         result = []
         for i, exp in enumerate(self.exps, start=1):
@@ -29,25 +36,3 @@ class BaseContextBuilder(BaseModel, ABC):
             result.append(f"{i}. " + EXP_TEMPLATE.format(req=exp.req, resp=exp.resp, score=score_val))
 
         return "\n".join(result)
-
-    @staticmethod
-    def replace_content_between_markers(text: str, start_marker: str, end_marker: str, new_content: str) -> str:
-        """Replace the content between `start_marker` and `end_marker` in the text with `new_content`.
-
-        Args:
-            text (str): The original text.
-            new_content (str): The new content to replace the old content.
-            start_marker (str): The marker indicating the start of the content to be replaced, such as '# Example'.
-            end_marker (str): The marker indicating the end of the content to be replaced, such as '# Instruction'.
-
-        Returns:
-            str: The text with the content replaced.
-        """
-
-        pattern = re.compile(f"({start_marker}\n)(.*?)(\n{end_marker})", re.DOTALL)
-
-        def replacement(match):
-            return f"{match.group(1)}{new_content}\n{match.group(3)}"
-
-        replaced_text = pattern.sub(replacement, text)
-        return replaced_text
diff --git a/metagpt/exp_pool/context_builders/role_zero.py b/metagpt/exp_pool/context_builders/role_zero.py
index e9ab83d90..b492ca5ca 100644
--- a/metagpt/exp_pool/context_builders/role_zero.py
+++ b/metagpt/exp_pool/context_builders/role_zero.py
@@ -1,15 +1,19 @@
 """RoleZero context builder."""
-import copy
-import json
+
+import re
 
 from metagpt.exp_pool.context_builders.base import BaseContextBuilder
 
 
 class RoleZeroContextBuilder(BaseContextBuilder):
-    async def build(self, *args, **kwargs) -> list[dict]:
+    async def build(self, **kwargs) -> list[dict]:
         """Builds the context by updating the req with formatted experiences.
 
-        If there are no experiences, retains the original examples in req, otherwise replaces the examples with the formatted experiences.
+        Args:
+            **kwargs: Arbitrary keyword arguments, expecting 'req' as a key.
+
+        Returns:
+            list[dict]: The updated request with formatted experiences or the original request if no experiences are available.
         """
 
         req = kwargs.get("req", [])
@@ -28,23 +32,23 @@ class RoleZeroContextBuilder(BaseContextBuilder):
         return self.replace_content_between_markers(text, "# Example", "# Instruction", new_example_content)
 
     @staticmethod
-    def req_serialize(req: list[dict]) -> str:
-        """Serialize the request for database storage, ensuring it is a string.
+    def replace_content_between_markers(text: str, start_marker: str, end_marker: str, new_content: str) -> str:
+        """Replace the content between `start_marker` and `end_marker` in the text with `new_content`.
 
-        This function deep copies the request and modifies the content of the last element
-        to remove unnecessary sections, making the request more concise.
+        Args:
+            text (str): The original text.
+            new_content (str): The new content to replace the old content.
+            start_marker (str): The marker indicating the start of the content to be replaced, such as '# Example'.
+            end_marker (str): The marker indicating the end of the content to be replaced, such as '# Instruction'.
+
+        Returns:
+            str: The text with the content replaced.
         """
 
-        req_copy = copy.deepcopy(req)
+        pattern = re.compile(f"({start_marker}\n)(.*?)(\n{end_marker})", re.DOTALL)
 
-        last_content = req_copy[-1]["content"]
-        last_content = RoleZeroContextBuilder.replace_content_between_markers(
-            last_content, "# Data Structure", "# Current Plan", ""
-        )
-        last_content = RoleZeroContextBuilder.replace_content_between_markers(
-            last_content, "# Example", "# Instruction", ""
-        )
+        def replacement(match):
+            return f"{match.group(1)}{new_content}\n{match.group(3)}"
 
-        req_copy[-1]["content"] = last_content
-
-        return json.dumps(req_copy)
+        replaced_text = pattern.sub(replacement, text)
+        return replaced_text
diff --git a/metagpt/exp_pool/context_builders/simple.py b/metagpt/exp_pool/context_builders/simple.py
index 35e2e1c8a..565855664 100644
--- a/metagpt/exp_pool/context_builders/simple.py
+++ b/metagpt/exp_pool/context_builders/simple.py
@@ -4,21 +4,21 @@
 from metagpt.exp_pool.context_builders.base import BaseContextBuilder
 
 SIMPLE_CONTEXT_TEMPLATE = """
-{req}
+## Context
 
 ### Experiences
 -----
 {exps}
 -----
 
+## User Requirement
+{req}
+
 ## Instruction
 Consider **Experiences** to generate a better answer.
 """
 
 
 class SimpleContextBuilder(BaseContextBuilder):
-    async def build(self, *args, **kwargs) -> str:
-        req = kwargs.get("req", "")
-        exps = self.format_exps()
-
-        return SIMPLE_CONTEXT_TEMPLATE.format(req=req, exps=exps) if exps else req
+    async def build(self, **kwargs) -> str:
+        return SIMPLE_CONTEXT_TEMPLATE.format(req=kwargs.get("req", ""), exps=self.format_exps())
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 62f766b9d..deb3faafc 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -1,6 +1,7 @@
 """Experience Decorator."""
 
 import asyncio
+import copy
 import functools
 from typing import Any, Callable, Optional, TypeVar
 
@@ -12,6 +13,7 @@ from metagpt.exp_pool.manager import ExperienceManager, exp_manager
 from metagpt.exp_pool.perfect_judges import BasePerfectJudge, SimplePerfectJudge
 from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
 from metagpt.exp_pool.scorers import BaseScorer, SimpleScorer
+from metagpt.exp_pool.serializers import BaseSerializer, SimpleSerializer
 from metagpt.logs import logger
 from metagpt.utils.async_helper import NestAsyncio
 from metagpt.utils.exceptions import handle_exception
@@ -26,9 +28,7 @@ def exp_cache(
     scorer: Optional[BaseScorer] = None,
     perfect_judge: Optional[BasePerfectJudge] = None,
     context_builder: Optional[BaseContextBuilder] = None,
-    req_serialize: Optional[Callable[..., str]] = None,
-    resp_serialize: Optional[Callable[..., str]] = None,
-    resp_deserialize: Optional[Callable[[str], Any]] = None,
+    serializer: Optional[BaseSerializer] = None,
     tag: Optional[str] = None,
 ):
     """Decorator to get a perfect experience, otherwise, it executes the function, and create a new experience.
@@ -44,9 +44,7 @@ def exp_cache(
         scorer: Evaluate experience. Default to `SimpleScorer()`.
         perfect_judge: Determines if an experience is perfect. Defaults to `SimplePerfectJudge()`.
         context_builder: Build the context from exps and the function parameters. Default to `SimpleContextBuilder()`.
-        req_serialize: Serializes the request for storage. Defaults to `lambda req: str(req)`.
-        resp_serialize: Serializes the function's return value for storage. Defaults to `lambda resp: str(resp)`.
-        resp_deserialize: Deserializes the stored response back to the function's return value. Defaults to `lambda resp: resp`.
+        serializer: Serializes the request and the function's return value for storage, deserializes the stored response back to the function's return value. Defaults to `SimpleSerializer()`.
         tag: An optional tag for the experience. Default to `ClassName.method_name` or `function_name`.
     """
 
@@ -65,9 +63,7 @@ def exp_cache(
                 exp_scorer=scorer,
                 exp_perfect_judge=perfect_judge,
                 context_builder=context_builder,
-                req_serialize=req_serialize,
-                resp_serialize=resp_serialize,
-                resp_deserialize=resp_deserialize,
+                serializer=serializer,
                 tag=tag,
             )
 
@@ -96,9 +92,7 @@ class ExpCacheHandler(BaseModel):
     exp_scorer: Optional[BaseScorer] = None
     exp_perfect_judge: Optional[BasePerfectJudge] = None
     context_builder: Optional[BaseContextBuilder] = None
-    req_serialize: Optional[Callable[..., str]] = None
-    resp_serialize: Optional[Callable[..., str]] = None
-    resp_deserialize: Optional[Callable[[str], Any]] = None
+    serializer: Optional[BaseSerializer] = None
     tag: Optional[str] = None
 
     _exps: list[Experience] = None
@@ -120,12 +114,10 @@ class ExpCacheHandler(BaseModel):
         self.exp_scorer = self.exp_scorer or SimpleScorer()
         self.exp_perfect_judge = self.exp_perfect_judge or SimplePerfectJudge()
         self.context_builder = self.context_builder or SimpleContextBuilder()
-        self.req_serialize = self.req_serialize or (lambda resp: str(resp))
-        self.resp_serialize = self.resp_serialize or (lambda resp: str(resp))
-        self.resp_deserialize = self.resp_deserialize or (lambda resp: resp)
+        self.serializer = self.serializer or SimpleSerializer()
         self.tag = self.tag or self._generate_tag()
 
-        self._req = self.req_serialize(self.kwargs["req"])
+        self._req = self.serializer.serialize_req(copy.deepcopy(self.kwargs["req"]))
 
         return self
 
@@ -140,7 +132,7 @@ class ExpCacheHandler(BaseModel):
         for exp in self._exps:
             if await self.exp_perfect_judge.is_perfect_exp(exp, self._req, *self.args, **self.kwargs):
                 logger.info(f"Get one perfect experience: {exp.req[:20]}...")
-                return self.resp_deserialize(exp.resp)
+                return self.serializer.deserialize_resp(exp.resp)
 
         return None
 
@@ -148,7 +140,7 @@ class ExpCacheHandler(BaseModel):
         """Execute the function, and save resp."""
 
         self._raw_resp = await self._execute_function()
-        self._resp = self.resp_serialize(self._raw_resp)
+        self._resp = self.serializer.serialize_resp(copy.deepcopy(self._raw_resp))
 
     @handle_exception
     async def process_experience(self):
@@ -204,7 +196,7 @@ class ExpCacheHandler(BaseModel):
     async def _build_context(self) -> str:
         self.context_builder.exps = self._exps
 
-        return await self.context_builder.build(*self.args, **self.kwargs)
+        return await self.context_builder.build(**self.kwargs)
 
     async def _execute_function(self):
         self.kwargs["req"] = await self._build_context()
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 23198eb02..649210a79 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -47,14 +47,12 @@ class ExperienceManager(BaseModel):
 
             self.storage = SimpleEngine.from_objs(retriever_configs=retriever_configs, ranker_configs=ranker_configs)
 
-        self.init_exp_pool()
-
         logger.debug(f"exp_pool config: {self.config.exp_pool}")
         return self
 
     @handle_exception
     def init_exp_pool(self):
-        if not self.config.exp_pool.init_exp:
+        if not self.config.exp_pool.enable_write:
             return
 
         if self._has_exps():
diff --git a/metagpt/exp_pool/serializers/__init__.py b/metagpt/exp_pool/serializers/__init__.py
new file mode 100644
index 000000000..8e1045588
--- /dev/null
+++ b/metagpt/exp_pool/serializers/__init__.py
@@ -0,0 +1,9 @@
+"""Serializers init."""
+
+from metagpt.exp_pool.serializers.base import BaseSerializer
+from metagpt.exp_pool.serializers.simple import SimpleSerializer
+from metagpt.exp_pool.serializers.action_node import ActionNodeSerializer
+from metagpt.exp_pool.serializers.role_zero import RoleZeroSerializer
+
+
+__all__ = ["BaseSerializer", "SimpleSerializer", "ActionNodeSerializer", "RoleZeroSerializer"]
diff --git a/metagpt/exp_pool/serializers/action_node.py b/metagpt/exp_pool/serializers/action_node.py
new file mode 100644
index 000000000..7746d6be4
--- /dev/null
+++ b/metagpt/exp_pool/serializers/action_node.py
@@ -0,0 +1,36 @@
+"""ActionNode Serializer."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Type
+
+# Import ActionNode only for type checking to avoid circular imports
+if TYPE_CHECKING:
+    from metagpt.actions.action_node import ActionNode
+
+from metagpt.exp_pool.serializers.simple import SimpleSerializer
+
+
+class ActionNodeSerializer(SimpleSerializer):
+    def serialize_resp(self, resp: ActionNode) -> str:
+        return resp.instruct_content.model_dump_json()
+
+    def deserialize_resp(self, resp: str) -> ActionNode:
+        """Customized deserialization, it will be triggered when a perfect experience is found.
+
+        ActionNode cannot be serialized, it throws an error 'cannot pickle 'SSLContext' object'.
+        """
+
+        class InstructContent:
+            def __init__(self, json_data):
+                self.json_data = json_data
+
+            def model_dump_json(self):
+                return self.json_data
+
+        from metagpt.actions.action_node import ActionNode
+
+        action_node = ActionNode(key="", expected_type=Type[str], instruction="", example="")
+        action_node.instruct_content = InstructContent(resp)
+
+        return action_node
diff --git a/metagpt/exp_pool/serializers/base.py b/metagpt/exp_pool/serializers/base.py
new file mode 100644
index 000000000..82a0ed8c4
--- /dev/null
+++ b/metagpt/exp_pool/serializers/base.py
@@ -0,0 +1,22 @@
+"""Base serializer."""
+
+from abc import ABC, abstractmethod
+from typing import Any
+
+from pydantic import BaseModel, ConfigDict
+
+
+class BaseSerializer(BaseModel, ABC):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    @abstractmethod
+    def serialize_req(self, req: Any) -> str:
+        """Serializes the request for storage."""
+
+    @abstractmethod
+    def serialize_resp(self, resp: Any) -> str:
+        """Serializes the function's return value for storage."""
+
+    @abstractmethod
+    def deserialize_resp(self, resp: str) -> Any:
+        """Deserializes the stored response back to the function's return value"""
diff --git a/metagpt/exp_pool/serializers/role_zero.py b/metagpt/exp_pool/serializers/role_zero.py
new file mode 100644
index 000000000..75e5d5ecb
--- /dev/null
+++ b/metagpt/exp_pool/serializers/role_zero.py
@@ -0,0 +1,40 @@
+"""RoleZero Serializer."""
+
+import json
+
+from metagpt.exp_pool.context_builders import RoleZeroContextBuilder
+from metagpt.exp_pool.serializers.simple import SimpleSerializer
+
+
+class RoleZeroSerializer(SimpleSerializer):
+    def serialize_req(self, req: list[dict]) -> str:
+        """Serialize the request for database storage, ensuring it is a string.
+
+        This function modifies the content of the last element in the request to remove unnecessary sections,
+        making the request more concise.
+
+        Args:
+            req (list[dict]): The request to be serialized. Example:
+                [
+                    {"role": "user", "content": "..."},
+                    {"role": "assistant", "content": "..."},
+                    {"role": "user", "content": "..."},
+                ]
+
+        Returns:
+            str: The serialized request as a JSON string.
+        """
+        if not req:
+            return ""
+
+        last_content = req[-1]["content"]
+        last_content = RoleZeroContextBuilder.replace_content_between_markers(
+            last_content, "# Data Structure", "# Current Plan", ""
+        )
+        last_content = RoleZeroContextBuilder.replace_content_between_markers(
+            last_content, "# Example", "# Instruction", ""
+        )
+
+        req[-1]["content"] = last_content
+
+        return json.dumps(req)
diff --git a/metagpt/exp_pool/serializers/simple.py b/metagpt/exp_pool/serializers/simple.py
new file mode 100644
index 000000000..32fe29c9f
--- /dev/null
+++ b/metagpt/exp_pool/serializers/simple.py
@@ -0,0 +1,22 @@
+"""Simple Serializer."""
+
+from typing import Any
+
+from metagpt.exp_pool.serializers.base import BaseSerializer
+
+
+class SimpleSerializer(BaseSerializer):
+    def serialize_req(self, req: Any) -> str:
+        """Just use `str` to convert the request object into a string."""
+
+        return str(req)
+
+    def serialize_resp(self, resp: Any) -> str:
+        """Just use `str` to convert the response object into a string."""
+
+        return str(resp)
+
+    def deserialize_resp(self, resp: str) -> Any:
+        """Just return the string response as it is."""
+
+        return resp
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 103a77911..59c58861f 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -12,6 +12,7 @@ from metagpt.actions import Action
 from metagpt.actions.di.run_command import RunCommand
 from metagpt.exp_pool import exp_cache
 from metagpt.exp_pool.context_builders import RoleZeroContextBuilder
+from metagpt.exp_pool.serializers import RoleZeroSerializer
 from metagpt.logs import logger
 from metagpt.prompts.di.role_zero import (
     CMD_PROMPT,
@@ -165,9 +166,7 @@ class RoleZero(Role):
 
         return True
 
-    @exp_cache(
-        context_builder=RoleZeroContextBuilder(), req_serialize=lambda req: RoleZeroContextBuilder.req_serialize(req)
-    )
+    @exp_cache(context_builder=RoleZeroContextBuilder(), serializer=RoleZeroSerializer())
     async def llm_cached_aask(self, *, req: list[dict], system_msgs: list[str]) -> str:
         return await self.llm.aask(req, system_msgs=system_msgs)
 
diff --git a/tests/metagpt/exp_pool/test_context_builders/test_base_context_builder.py b/tests/metagpt/exp_pool/test_context_builders/test_base_context_builder.py
index 17696e1b4..0a160fb42 100644
--- a/tests/metagpt/exp_pool/test_context_builders/test_base_context_builder.py
+++ b/tests/metagpt/exp_pool/test_context_builders/test_base_context_builder.py
@@ -30,16 +30,3 @@ class TestBaseContextBuilder:
             ]
         )
         assert result == expected
-
-    def test_replace_content_between_markers(self):
-        text = "Start\n# Example\nOld content\n# Instruction\nEnd"
-        new_content = "New content"
-        result = BaseContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
-        expected = "Start\n# Example\nNew content\n\n# Instruction\nEnd"
-        assert result == expected
-
-    def test_replace_content_between_markers_no_match(self):
-        text = "Start\nNo markers\nEnd"
-        new_content = "New content"
-        result = BaseContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
-        assert result == text
diff --git a/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py b/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
index 0ea04432d..611d68211 100644
--- a/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
+++ b/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
@@ -30,9 +30,22 @@ class TestRoleZeroContextBuilder:
         assert result == [{"content": "Updated content"}]
 
     def test_replace_example_content(self, context_builder, mocker):
-        mocker.patch.object(BaseContextBuilder, "replace_content_between_markers", return_value="Replaced content")
+        mocker.patch.object(RoleZeroContextBuilder, "replace_content_between_markers", return_value="Replaced content")
         result = context_builder.replace_example_content("Original text", "New example content")
         assert result == "Replaced content"
         context_builder.replace_content_between_markers.assert_called_once_with(
             "Original text", "# Example", "# Instruction", "New example content"
         )
+
+    def test_replace_content_between_markers(self):
+        text = "Start\n# Example\nOld content\n# Instruction\nEnd"
+        new_content = "New content"
+        result = RoleZeroContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
+        expected = "Start\n# Example\nNew content\n\n# Instruction\nEnd"
+        assert result == expected
+
+    def test_replace_content_between_markers_no_match(self):
+        text = "Start\nNo markers\nEnd"
+        new_content = "New content"
+        result = RoleZeroContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
+        assert result == text
diff --git a/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py b/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py
index e96addab9..b6d0f642e 100644
--- a/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py
+++ b/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py
@@ -32,7 +32,8 @@ class TestSimpleContextBuilder:
         req = "Test request"
         result = await context_builder.build(req=req)
 
-        assert result == req
+        expected = SIMPLE_CONTEXT_TEMPLATE.format(req=req, exps="")
+        assert result == expected
 
     @pytest.mark.asyncio
     async def test_build_without_req(self, context_builder, mocker):
diff --git a/tests/metagpt/exp_pool/test_manager.py b/tests/metagpt/exp_pool/test_manager.py
index c12fc7e8c..6d0693efd 100644
--- a/tests/metagpt/exp_pool/test_manager.py
+++ b/tests/metagpt/exp_pool/test_manager.py
@@ -11,9 +11,7 @@ from metagpt.rag.engines import SimpleEngine
 class TestExperienceManager:
     @pytest.fixture
     def mock_config(self):
-        return Config(
-            llm=LLMConfig(), exp_pool=ExperiencePoolConfig(enable_write=True, enable_read=True, init_exp=False)
-        )
+        return Config(llm=LLMConfig(), exp_pool=ExperiencePoolConfig(enable_write=True, enable_read=True))
 
     @pytest.fixture
     def mock_storage(self, mocker):
@@ -65,30 +63,6 @@ class TestExperienceManager:
         result = await mock_experience_manager.query_exps("query")
         assert result == []
 
-    def test_init_exp_pool(self, mock_experience_manager, mock_config, mocker):
-        mock_experience_manager._has_exps = mocker.MagicMock(return_value=False)
-        mock_experience_manager._init_teamleader_exps = mocker.MagicMock()
-        mock_experience_manager._init_engineer2_exps = mocker.MagicMock()
-
-        mock_config.exp_pool.init_exp = True
-        mock_experience_manager.init_exp_pool()
-
-        mock_experience_manager._has_exps.assert_called_once()
-        mock_experience_manager._init_teamleader_exps.assert_called_once()
-        mock_experience_manager._init_engineer2_exps.assert_called_once()
-
-    def test_init_exp_pool_already_has_exps(self, mock_experience_manager, mock_config, mocker):
-        mock_experience_manager._has_exps = mocker.MagicMock(return_value=True)
-        mock_experience_manager._init_teamleader_exps = mocker.MagicMock()
-        mock_experience_manager._init_engineer2_exps = mocker.MagicMock()
-
-        mock_config.exp_pool.init_exp = True
-        mock_experience_manager.init_exp_pool()
-
-        mock_experience_manager._has_exps.assert_called_once()
-        mock_experience_manager._init_teamleader_exps.assert_not_called()
-        mock_experience_manager._init_engineer2_exps.assert_not_called()
-
     def test_has_exps(self, mock_experience_manager, mock_storage):
         mock_storage._retriever._vector_store._get.return_value.ids = ["id1"]
 

From 866d93b7dbaa331536a1c68c9637106a73dfae84 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Thu, 11 Jul 2024 10:26:42 +0800
Subject: [PATCH 39/51] when enable_write is false, skip evaluating and saving
 the experience

---
 metagpt/exp_pool/context_builders/role_zero.py |  7 +++----
 metagpt/exp_pool/decorator.py                  | 18 +++++++++++-------
 metagpt/exp_pool/serializers/base.py           | 11 +++++++++--
 metagpt/exp_pool/serializers/role_zero.py      |  8 +++++---
 metagpt/strategy/experience_retriever.py       |  4 ++--
 5 files changed, 30 insertions(+), 18 deletions(-)

diff --git a/metagpt/exp_pool/context_builders/role_zero.py b/metagpt/exp_pool/context_builders/role_zero.py
index b492ca5ca..6407314ac 100644
--- a/metagpt/exp_pool/context_builders/role_zero.py
+++ b/metagpt/exp_pool/context_builders/role_zero.py
@@ -15,16 +15,15 @@ class RoleZeroContextBuilder(BaseContextBuilder):
         Returns:
             list[dict]: The updated request with formatted experiences or the original request if no experiences are available.
         """
-
         req = kwargs.get("req", [])
         if not req:
             return req
 
-        exps_str = self.format_exps()
-        if not exps_str:
+        exps = self.format_exps()
+        if not exps:
             return req
 
-        req[-1]["content"] = self.replace_example_content(req[-1].get("content", ""), exps_str)
+        req[-1]["content"] = self.replace_example_content(req[-1].get("content", ""), exps)
 
         return req
 
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index deb3faafc..0a9a83818 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -1,7 +1,6 @@
 """Experience Decorator."""
 
 import asyncio
-import copy
 import functools
 from typing import Any, Callable, Optional, TypeVar
 
@@ -33,9 +32,11 @@ def exp_cache(
 ):
     """Decorator to get a perfect experience, otherwise, it executes the function, and create a new experience.
 
-    1. This can be applied to both synchronous and asynchronous functions.
-    2. The function must have a `req` parameter, and it must be provided as a keyword argument.
-    3. If `config.exp_pool.enable_read` is False, the decorator will just directly execute the function.
+    Note:
+        1. This can be applied to both synchronous and asynchronous functions.
+        2. The function must have a `req` parameter, and it must be provided as a keyword argument.
+        3. If `config.exp_pool.enable_read` is False, the decorator will just directly execute the function.
+        4. If `config.exp_pool.enable_write` is False, the decorator will skip evaluating and saving the experience.
 
     Args:
         _func: Just to make the decorator more flexible, for example, it can be used directly with @exp_cache by default, without the need for @exp_cache().
@@ -68,11 +69,14 @@ def exp_cache(
             )
 
             await handler.fetch_experiences()
+
             if exp := await handler.get_one_perfect_exp():
                 return exp
 
             await handler.execute_function()
-            await handler.process_experience()
+
+            if config.exp_pool.enable_write:
+                await handler.process_experience()
 
             return handler._raw_resp
 
@@ -117,7 +121,7 @@ class ExpCacheHandler(BaseModel):
         self.serializer = self.serializer or SimpleSerializer()
         self.tag = self.tag or self._generate_tag()
 
-        self._req = self.serializer.serialize_req(copy.deepcopy(self.kwargs["req"]))
+        self._req = self.serializer.serialize_req(self.kwargs["req"])
 
         return self
 
@@ -140,7 +144,7 @@ class ExpCacheHandler(BaseModel):
         """Execute the function, and save resp."""
 
         self._raw_resp = await self._execute_function()
-        self._resp = self.serializer.serialize_resp(copy.deepcopy(self._raw_resp))
+        self._resp = self.serializer.serialize_resp(self._raw_resp)
 
     @handle_exception
     async def process_experience(self):
diff --git a/metagpt/exp_pool/serializers/base.py b/metagpt/exp_pool/serializers/base.py
index 82a0ed8c4..9d00a05b2 100644
--- a/metagpt/exp_pool/serializers/base.py
+++ b/metagpt/exp_pool/serializers/base.py
@@ -11,11 +11,18 @@ class BaseSerializer(BaseModel, ABC):
 
     @abstractmethod
     def serialize_req(self, req: Any) -> str:
-        """Serializes the request for storage."""
+        """Serializes the request for storage.
+
+        Do not modify req. If modification is necessary, use copy.deepcopy to create a copy first.
+        Note that copy.deepcopy may raise errors, such as TypeError: cannot pickle '_thread.RLock' object.
+        """
 
     @abstractmethod
     def serialize_resp(self, resp: Any) -> str:
-        """Serializes the function's return value for storage."""
+        """Serializes the function's return value for storage.
+
+        Do not modify resp. The rest is the same as `serialize_req`.
+        """
 
     @abstractmethod
     def deserialize_resp(self, resp: str) -> Any:
diff --git a/metagpt/exp_pool/serializers/role_zero.py b/metagpt/exp_pool/serializers/role_zero.py
index 75e5d5ecb..7876ef12a 100644
--- a/metagpt/exp_pool/serializers/role_zero.py
+++ b/metagpt/exp_pool/serializers/role_zero.py
@@ -1,5 +1,6 @@
 """RoleZero Serializer."""
 
+import copy
 import json
 
 from metagpt.exp_pool.context_builders import RoleZeroContextBuilder
@@ -27,7 +28,8 @@ class RoleZeroSerializer(SimpleSerializer):
         if not req:
             return ""
 
-        last_content = req[-1]["content"]
+        req_copy = copy.deepcopy(req)
+        last_content = req_copy[-1]["content"]
         last_content = RoleZeroContextBuilder.replace_content_between_markers(
             last_content, "# Data Structure", "# Current Plan", ""
         )
@@ -35,6 +37,6 @@ class RoleZeroSerializer(SimpleSerializer):
             last_content, "# Example", "# Instruction", ""
         )
 
-        req[-1]["content"] = last_content
+        req_copy[-1]["content"] = last_content
 
-        return json.dumps(req)
+        return json.dumps(req_copy)
diff --git a/metagpt/strategy/experience_retriever.py b/metagpt/strategy/experience_retriever.py
index 7bcd4be11..32f5c2316 100644
--- a/metagpt/strategy/experience_retriever.py
+++ b/metagpt/strategy/experience_retriever.py
@@ -798,13 +798,13 @@ Explanation: I will first need to read the system design document and the projec
     {
         "command_name": "Editor.read",
         "args": {
-            "path": "/tmp/docs/project_schedule.json"
+            "path": "/tmp/project_schedule.json"
         }
     },
     {
         "command_name": "Editor.read",
         "args": {
-            "path": "/tmp/docs/system_design.json"
+            "path": "/tmp/system_design.json"
         }
     }
 ]

From fb4446c0a963918598296e7095b375088ea19acd Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Thu, 11 Jul 2024 18:01:35 +0800
Subject: [PATCH 40/51] update serialize_req to make the request more concise

---
 metagpt/exp_pool/serializers/role_zero.py | 36 ++++++++++++++++++-----
 1 file changed, 29 insertions(+), 7 deletions(-)

diff --git a/metagpt/exp_pool/serializers/role_zero.py b/metagpt/exp_pool/serializers/role_zero.py
index 7876ef12a..82a32e36b 100644
--- a/metagpt/exp_pool/serializers/role_zero.py
+++ b/metagpt/exp_pool/serializers/role_zero.py
@@ -11,8 +11,7 @@ class RoleZeroSerializer(SimpleSerializer):
     def serialize_req(self, req: list[dict]) -> str:
         """Serialize the request for database storage, ensuring it is a string.
 
-        This function modifies the content of the last element in the request to remove unnecessary sections,
-        making the request more concise.
+        This function does not modify `req`; it only extracts the necessary content from `req` because `req` may be very lengthy and could cause embedding errors.
 
         Args:
             req (list[dict]): The request to be serialized. Example:
@@ -28,8 +27,33 @@ class RoleZeroSerializer(SimpleSerializer):
         if not req:
             return ""
 
-        req_copy = copy.deepcopy(req)
-        last_content = req_copy[-1]["content"]
+        filtered_req = self._filter_req(req)
+        self._clean_last_entry_content(filtered_req)
+
+        return json.dumps(filtered_req)
+
+    def _filter_req(self, req: list[dict]) -> list[dict]:
+        """Filter the request to include only necessary items and the last entry.
+
+        Args:
+            req (list[dict]): The original request.
+
+        Returns:
+            list[dict]: The filtered request.
+        """
+
+        filtered_req = [
+            copy.deepcopy(item) for item in req if "Command Editor.read executed: file_path" in item["content"]
+        ]
+        filtered_req.append(copy.deepcopy(req[-1]))
+
+        return filtered_req
+
+    def _clean_last_entry_content(self, req: list[dict]):
+        """Modifies the content of the last element in the request to remove unnecessary sections, making the request more concise."""
+
+        last_content = req[-1]["content"]
+
         last_content = RoleZeroContextBuilder.replace_content_between_markers(
             last_content, "# Data Structure", "# Current Plan", ""
         )
@@ -37,6 +61,4 @@ class RoleZeroSerializer(SimpleSerializer):
             last_content, "# Example", "# Instruction", ""
         )
 
-        req_copy[-1]["content"] = last_content
-
-        return json.dumps(req_copy)
+        req[-1]["content"] = last_content

From 39360b41c5c69c2b926e56562a5d627b28afb000 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Thu, 11 Jul 2024 22:14:04 +0800
Subject: [PATCH 41/51] init exp pool

---
 .../exp_pool/init_exp_pool/engineer_exps.py   | 115 +++++++++++
 examples/exp_pool/init_exp_pool/main.py       |  55 ++++++
 .../init_exp_pool/team_leader_exps.py         | 181 ++++++++++++++++++
 metagpt/exp_pool/manager.py                   |  40 +---
 tests/metagpt/exp_pool/test_manager.py        |  18 --
 .../test_serializers/test_action_node.py      |  35 ++++
 .../test_serializers/test_role_zero.py        |  77 ++++++++
 .../exp_pool/test_serializers/test_simple.py  |  44 +++++
 8 files changed, 512 insertions(+), 53 deletions(-)
 create mode 100644 examples/exp_pool/init_exp_pool/engineer_exps.py
 create mode 100644 examples/exp_pool/init_exp_pool/main.py
 create mode 100644 examples/exp_pool/init_exp_pool/team_leader_exps.py
 create mode 100644 tests/metagpt/exp_pool/test_serializers/test_action_node.py
 create mode 100644 tests/metagpt/exp_pool/test_serializers/test_role_zero.py
 create mode 100644 tests/metagpt/exp_pool/test_serializers/test_simple.py

diff --git a/examples/exp_pool/init_exp_pool/engineer_exps.py b/examples/exp_pool/init_exp_pool/engineer_exps.py
new file mode 100644
index 000000000..e494fdaa3
--- /dev/null
+++ b/examples/exp_pool/init_exp_pool/engineer_exps.py
@@ -0,0 +1,115 @@
+import json
+
+OBJ1 = [
+    {
+        "role": "user",
+        "content": '\n# Data Structure\n\n\n# Current Plan\n{\'goal\': "Please write a 1024 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at \'/tmp/system_design.json\' and the project schedule at \'/tmp/project_schedule.json\' for detailed information.", \'tasks\': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON\'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named "src". If you know the project path, then write in a "src" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW "Data structures and interfaces". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON\'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON\'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON\'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        "command_name": str,\n        "args": {"arg_name": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n',
+    }
+]
+OBJ2 = [
+    {
+        "role": "user",
+        "content": "Command Editor.read executed: file_path='/tmp/system_design.json' block_content=\"...\"\n\nCommand Editor.read executed: file_path='/tmp/project_schedule.json' block_content=\"...\"",
+    },
+    {
+        "role": "user",
+        "content": '\n# Data Structure\n\n\n# Current Plan\n{\'goal\': "Please write a 1024 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at \'/tmp/system_design.json\' and the project schedule at \'/tmp/project_schedule.json\' for detailed information.", \'tasks\': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON\'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named "src". If you know the project path, then write in a "src" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW "Data structures and interfaces". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON\'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON\'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON\'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        "command_name": str,\n        "args": {"arg_name": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n',
+    },
+]
+
+REQ1 = json.dumps(OBJ1)
+REQ2 = json.dumps(OBJ2)
+
+RESP1 = """
+Explanation: To begin the task of creating a 2048 game using JavaScript and HTML, I need to first read the system design and project schedule documents to understand the specific requirements and architecture outlined for the game development.
+
+```json
+[
+    {
+        "command_name": "Editor.read",
+        "args": {
+            "path": "/tmp/system_design.json"
+        }
+    },
+    {
+        "command_name": "Editor.read",
+        "args": {
+            "path": "/tmp/project_schedule.json"
+        }
+    }
+]
+```
+"""
+RESP2 = """
+Explanation: Based on the system design and project schedule, I will create a plan to implement the 2048 game. The plan will consist of tasks to create the necessary files (`index.html`, `style.css`, and `script.js`) and review each file to ensure they meet the design specifications.
+
+Here's the plan:
+
+1. **Task 1**: Create `index.html` - This file will contain the HTML structure necessary for the game's UI.
+2. **Task 2**: Create `style.css` - This file will define the CSS styles to make the game visually appealing and responsive.
+3. **Task 3**: Create `script.js` - This file will contain the JavaScript code for the game logic and UI interactions.
+4. **Code Review Tasks**: Review each file to ensure they meet the project requirements and adhere to the system design.
+
+Let's start by appending the first task to the plan.
+
+```json
+[
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "1",
+            "dependent_task_ids": [],
+            "instruction": "Create the index.html file with the basic HTML structure for the 2048 game.",
+            "assignee": "Alex"
+        }
+    },
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "2",
+            "dependent_task_ids": ["1"],
+            "instruction": "Create the style.css file with the necessary CSS to style the 2048 game.",
+            "assignee": "Alex"
+        }
+    },
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "3",
+            "dependent_task_ids": ["1", "2"],
+            "instruction": "Create the script.js file containing the JavaScript logic for the 2048 game.",
+            "assignee": "Alex"
+        }
+    },
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "4",
+            "dependent_task_ids": ["1"],
+            "instruction": "Use ReviewAndRewriteCode to review the code in index.html to ensure it meets the design specifications.",
+            "assignee": "Alex"
+        }
+    },
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "5",
+            "dependent_task_ids": ["2"],
+            "instruction": "Use ReviewAndRewriteCode to review the code in style.css to ensure it meets the design specifications.",
+            "assignee": "Alex"
+        }
+    },
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "6",
+            "dependent_task_ids": ["3"],
+            "instruction": "Use ReviewAndRewriteCode to review the code in script.js to ensure it meets the design specifications.",
+            "assignee": "Alex"
+        }
+    }
+]
+```
+"""
+
+ENGINEER_EXPERIENCES = ((REQ1, RESP1), (REQ2, RESP2))
diff --git a/examples/exp_pool/init_exp_pool/main.py b/examples/exp_pool/init_exp_pool/main.py
new file mode 100644
index 000000000..849834513
--- /dev/null
+++ b/examples/exp_pool/init_exp_pool/main.py
@@ -0,0 +1,55 @@
+"""Init experience pool.
+
+Put some useful experiences into the experience pool.
+"""
+
+import asyncio
+
+from examples.exp_pool.init_exp_pool.engineer_exps import ENGINEER_EXPERIENCES
+from examples.exp_pool.init_exp_pool.team_leader_exps import TEAM_LEADER_EXPERIENCES
+from metagpt.exp_pool import exp_manager
+from metagpt.exp_pool.schema import EntryType, Experience, Metric, Score
+from metagpt.logs import logger
+
+
+async def add_exp(req: str, resp: str, tag: str, metric: Metric = None):
+    exp = Experience(
+        req=req,
+        resp=resp,
+        entry_type=EntryType.MANUAL,
+        tag=tag,
+        metric=metric or Metric(score=Score(val=10, reason="Manual")),
+    )
+
+    exp_manager.config.exp_pool.enable_write = True
+    exp_manager.create_exp(exp)
+    logger.info(f"New experience created for the request `{req[:10]}`.")
+
+
+async def add_teamleader_exps():
+    tag = "TeamLeader.llm_cached_aask"
+
+    for req, resp in TEAM_LEADER_EXPERIENCES:
+        await add_exp(req=req, resp=resp, tag=tag)
+
+
+async def add_engineer_exps():
+    tag = "Engineer2.llm_cached_aask"
+
+    for req, resp in ENGINEER_EXPERIENCES:
+        await add_exp(req=req, resp=resp, tag=tag)
+
+
+def query_exps_count():
+    count = exp_manager.get_exps_count()
+    logger.info(f"Experiences Count: {count}")
+
+
+async def main():
+    await add_teamleader_exps()
+    await add_engineer_exps()
+    query_exps_count()
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/examples/exp_pool/init_exp_pool/team_leader_exps.py b/examples/exp_pool/init_exp_pool/team_leader_exps.py
new file mode 100644
index 000000000..da0bc3d81
--- /dev/null
+++ b/examples/exp_pool/init_exp_pool/team_leader_exps.py
@@ -0,0 +1,181 @@
+import json
+
+OBJ1 = [
+    {
+        "role": "user",
+        "content": "\n# Data Structure\n\n\n# Current Plan\n{'goal': \"from  to {'<all>'}: Create a cli snake game using Python.\", 'tasks': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        \"command_name\": str,\n        \"args\": {\"arg_name\": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n",
+    }
+]
+OBJ2 = [
+    {
+        "role": "user",
+        "content": "\n# Data Structure\n\n\n# Current Plan\n{'goal': \"from  to {'<all>'}: Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.\", 'tasks': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        \"command_name\": str,\n        \"args\": {\"arg_name\": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n",
+    }
+]
+OBJ4 = [
+    {
+        "role": "user",
+        "content": "\n# Data Structure\n\n\n# Current Plan\n{'goal': \"from  to {'<all>'}: how does the project go?\", 'tasks': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        \"command_name\": str,\n        \"args\": {\"arg_name\": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n",
+    }
+]
+
+REQ1 = json.dumps(OBJ1)
+REQ2 = json.dumps(OBJ2)
+REQ3 = """Conversation History:
+[
+    ...,
+    {'role': 'assistant', 'content': 'from Alice(Product Manager) to {'Bob'}: {'docs': {'20240424153821.json': {'root_path': 'docs/prd', 'filename': '20240424153821.json', 'content': '{"Language":"en_us","Programming Language":"Python","Original Requirements":"create a cli snake game","Project Name":"snake_game","Product Goals":["Develop an intuitive and addictive snake game",...], ...}}}}},
+]
+"""
+REQ4 = json.dumps(OBJ4)
+
+RESP1 = """
+Explanation: The requirement is about software development. Assign each tasks to a different team member based on their expertise. When publishing message to Product Manager, we copy original user requirement directly to ensure no information loss.
+```json
+[
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "1",
+            "dependent_task_ids": [],
+            "instruction": "Create a product requirement document (PRD) outlining the features, user interface, and user experience of the CLI python snake game.",
+            "assignee": "Alice"
+        }
+    },
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "2",
+            "dependent_task_ids": ["1"],
+            "instruction": "Design the software architecture for the CLI snake game, including the choice of programming language, libraries, and data flow.",
+            "assignee": "Bob"
+        }
+    },
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "3",
+            "dependent_task_ids": ["2"],
+            "instruction": "Break down the architecture into manageable tasks, identify task dependencies, and prepare a detailed task list for implementation.",
+            "assignee": "Eve"
+        }
+    },
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "4",
+            "dependent_task_ids": ["3"],
+            "instruction": "Implement the core game logic for the CLI snake game, including snake movement, food generation, and score tracking.",
+            "assignee": "Alex"
+        }
+    },
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "5",
+            "dependent_task_ids": ["4"],
+            "instruction": "Write comprehensive tests for the game logic and user interface to ensure functionality and reliability.",
+            "assignee": "Edward"
+        }
+    },
+    {
+        "command_name": "TeamLeader.publish_message",
+        "args": {
+            "content": "Create a cli snake game using Python",
+            "send_to": "Alice"
+        }
+    },
+    {
+        "command_name": "RoleZero.reply_to_human",
+        "args": {
+            "content": "I have assigned the tasks to the team members. Alice will create the PRD, Bob will design the software architecture, Eve will break down the architecture into tasks, Alex will implement the core game logic, and Edward will write comprehensive tests. The team will work on the project accordingly",
+        }
+    },
+    {
+        "command_name": "end"
+    }
+]
+```
+"""
+RESP2 = """
+Explanation: DON'T decompose requirement if it is a DATA-RELATED task, assign a single task directly to Data Analyst David. He will manage the decomposition and implementation.
+```json
+[
+    {
+        "command_name": "Plan.append_task",
+        "args": {
+            "task_id": "1",
+            "dependent_task_ids": [],
+            "instruction": "Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.",
+            "assignee": "David"
+        }
+    },
+    {
+        "command_name": "TeamLeader.publish_message",
+        "args": {
+            "content": "Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.",
+            "send_to": "David"
+        }
+    },
+    {
+        "command_name": "RoleZero.reply_to_human",
+        "args": {
+            "content": "I have assigned the task to David. He will break down the task further by himself and starts solving it.",
+        }
+    },
+    {
+        "command_name": "end"
+    }
+]
+```
+"""
+RESP3 = """
+Explanation: You received a message from Alice, the Product Manager, that she has completed the PRD, use Plan.finish_current_task to mark her task as finished and moves the plan to the next task. Based on plan status, next task is for Bob (Architect), publish a message asking him to start. The message content should contain important path info.
+```json
+[
+    {
+        "command_name": "Plan.finish_current_task",
+        "args": {}
+    },
+    {
+        "command_name": "TeamLeader.publish_message",
+            "args": {
+                "content": "Please design the software architecture for the snake game based on the PRD created by Alice. The PRD is at 'docs/prd/20240424153821.json'. Include the choice of programming language, libraries, and data flow, etc.",
+                "send_to": "Bob"
+            }
+    },
+    {
+        "command_name": "RoleZero.reply_to_human",
+        "args": {
+            "content": "Alice has completed the PRD. I have marked her task as finished and sent the PRD to Bob. Bob will work on the software architecture.",
+        }
+    },
+    {
+        "command_name": "end"
+    }
+]
+```
+"""
+RESP4 = """
+Explanation: The user is asking for a general update on the project status. Give a straight answer about the current task the team is working on and provide a summary of the completed tasks.
+```json
+[
+    {
+        "command_name": "RoleZero.reply_to_human",
+        "args": {
+            "content": "The team is currently working on ... We have completed ...",
+        }
+    },
+    {
+        "command_name": "end"
+    }
+]
+```
+"""
+
+TEAM_LEADER_EXPERIENCES = (
+    (REQ1, RESP1),
+    (REQ2, RESP2),
+    (REQ3, RESP3),
+    (REQ4, RESP4),
+)
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index 649210a79..ba1a8bcf0 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -7,16 +7,12 @@ from metagpt.config2 import Config, config
 from metagpt.exp_pool.schema import (
     DEFAULT_COLLECTION_NAME,
     DEFAULT_SIMILARITY_TOP_K,
-    EntryType,
     Experience,
-    Metric,
     QueryType,
-    Score,
 )
 from metagpt.logs import logger
 from metagpt.rag.engines import SimpleEngine
 from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
-from metagpt.strategy.experience_retriever import ENGINEER_EXAMPLE, TL_EXAMPLE
 from metagpt.utils.exceptions import handle_exception
 
 
@@ -50,17 +46,9 @@ class ExperienceManager(BaseModel):
         logger.debug(f"exp_pool config: {self.config.exp_pool}")
         return self
 
-    @handle_exception
-    def init_exp_pool(self):
-        if not self.config.exp_pool.enable_write:
-            return
-
-        if self._has_exps():
-            return
-
-        self._init_teamleader_exps()
-        self._init_engineer2_exps()
-        logger.info("`init_exp_pool` done.")
+    @property
+    def vector_store(self) -> ChromaVectorStore:
+        return self.storage._retriever._vector_store
 
     @handle_exception
     def create_exp(self, exp: Experience):
@@ -101,26 +89,8 @@ class ExperienceManager(BaseModel):
 
         return exps
 
-    def _has_exps(self) -> bool:
-        vector_store: ChromaVectorStore = self.storage._retriever._vector_store
-
-        return bool(vector_store._get(limit=1, where={}).ids)
-
-    def _init_exp(self, req: str, resp: str, tag: str, metric: Metric = None):
-        exp = Experience(
-            req=req,
-            resp=resp,
-            entry_type=EntryType.MANUAL,
-            tag=tag,
-            metric=metric or Metric(score=Score(val=9, reason="Manual")),
-        )
-        self.create_exp(exp)
-
-    def _init_teamleader_exps(self):
-        self._init_exp(req=TL_EXAMPLE, resp=TL_EXAMPLE, tag="TeamLeader.llm_cached_aask")
-
-    def _init_engineer2_exps(self):
-        self._init_exp(req=ENGINEER_EXAMPLE, resp=ENGINEER_EXAMPLE, tag="Engineer2.llm_cached_aask")
+    def get_exps_count(self) -> int:
+        return self.vector_store._collection.count()
 
 
 exp_manager = ExperienceManager()
diff --git a/tests/metagpt/exp_pool/test_manager.py b/tests/metagpt/exp_pool/test_manager.py
index 6d0693efd..2f712ad44 100644
--- a/tests/metagpt/exp_pool/test_manager.py
+++ b/tests/metagpt/exp_pool/test_manager.py
@@ -62,21 +62,3 @@ class TestExperienceManager:
         mock_config.exp_pool.enable_read = False
         result = await mock_experience_manager.query_exps("query")
         assert result == []
-
-    def test_has_exps(self, mock_experience_manager, mock_storage):
-        mock_storage._retriever._vector_store._get.return_value.ids = ["id1"]
-
-        assert mock_experience_manager._has_exps() is True
-
-        mock_storage._retriever._vector_store._get.return_value.ids = []
-        assert mock_experience_manager._has_exps() is False
-
-    def test_init_teamleader_exps(self, mock_experience_manager, mocker):
-        mock_experience_manager._init_exp = mocker.MagicMock()
-        mock_experience_manager._init_teamleader_exps()
-        mock_experience_manager._init_exp.assert_called_once()
-
-    def test_init_engineer2_exps(self, mock_experience_manager, mocker):
-        mock_experience_manager._init_exp = mocker.MagicMock()
-        mock_experience_manager._init_engineer2_exps()
-        mock_experience_manager._init_exp.assert_called_once()
diff --git a/tests/metagpt/exp_pool/test_serializers/test_action_node.py b/tests/metagpt/exp_pool/test_serializers/test_action_node.py
new file mode 100644
index 000000000..e4ab4684d
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_serializers/test_action_node.py
@@ -0,0 +1,35 @@
+from typing import Type
+
+import pytest
+
+from metagpt.actions.action_node import ActionNode
+from metagpt.exp_pool.serializers.action_node import ActionNodeSerializer
+
+
+class TestActionNodeSerializer:
+    @pytest.fixture
+    def serializer(self):
+        return ActionNodeSerializer()
+
+    @pytest.fixture
+    def action_node(self):
+        class InstructContent:
+            def __init__(self, json_data):
+                self.json_data = json_data
+
+            def model_dump_json(self):
+                return self.json_data
+
+        action_node = ActionNode(key="", expected_type=Type[str], instruction="", example="")
+        action_node.instruct_content = InstructContent('{"key": "value"}')
+
+        return action_node
+
+    def test_serialize_resp(self, serializer: ActionNodeSerializer, action_node: ActionNode):
+        serialized = serializer.serialize_resp(action_node)
+        assert serialized == '{"key": "value"}'
+
+    def test_deserialize_resp(self, serializer: ActionNodeSerializer):
+        deserialized = serializer.deserialize_resp('{"key": "value"}')
+        assert isinstance(deserialized, ActionNode)
+        assert deserialized.instruct_content.model_dump_json() == '{"key": "value"}'
diff --git a/tests/metagpt/exp_pool/test_serializers/test_role_zero.py b/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
new file mode 100644
index 000000000..4c1f3daf3
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
@@ -0,0 +1,77 @@
+import json
+
+import pytest
+
+from metagpt.exp_pool.serializers import RoleZeroSerializer
+
+
+class TestRoleZeroSerializer:
+    @pytest.fixture
+    def serializer(self):
+        return RoleZeroSerializer()
+
+    def test_serialize_req_empty_input(self, serializer: RoleZeroSerializer):
+        assert serializer.serialize_req([]) == ""
+
+    def test_serialize_req_with_content(self, serializer: RoleZeroSerializer):
+        req = [
+            {"content": "Command Editor.read executed: file_path=test.py"},
+            {"content": "Some other content"},
+            {
+                "content": "# Data Structure\nsome data\n# Current Plan\nsome plan\n# Example\nsome example\n# Instruction\nsome instruction"
+            },
+        ]
+        expected_output = json.dumps(
+            [
+                {"content": "Command Editor.read executed: file_path=test.py"},
+                {
+                    "content": "# Data Structure\n\n\n# Current Plan\nsome plan\n# Example\n\n\n# Instruction\nsome instruction"
+                },
+            ]
+        )
+        assert serializer.serialize_req(req) == expected_output
+
+    def test_filter_req(self, serializer: RoleZeroSerializer):
+        req = [
+            {"content": "Command Editor.read executed: file_path=test1.py"},
+            {"content": "Some other content"},
+            {"content": "Command Editor.read executed: file_path=test2.py"},
+            {"content": "Final content"},
+        ]
+        filtered_req = serializer._filter_req(req)
+        assert len(filtered_req) == 3
+        assert filtered_req[0]["content"] == "Command Editor.read executed: file_path=test1.py"
+        assert filtered_req[1]["content"] == "Command Editor.read executed: file_path=test2.py"
+        assert filtered_req[2]["content"] == "Final content"
+
+    def test_clean_last_entry_content(self, serializer: RoleZeroSerializer):
+        req = [
+            {"content": "Some content"},
+            {
+                "content": "# Data Structure\nsome data\n# Current Plan\nsome plan\n# Example\nsome example\n# Instruction\nsome instruction"
+            },
+        ]
+        serializer._clean_last_entry_content(req)
+        expected_content = (
+            "# Data Structure\n\n\n# Current Plan\nsome plan\n# Example\n\n\n# Instruction\nsome instruction"
+        )
+        assert req[-1]["content"] == expected_content
+
+    def test_integration(self, serializer: RoleZeroSerializer):
+        req = [
+            {"content": "Command Editor.read executed: file_path=test.py"},
+            {"content": "Some other content"},
+            {
+                "content": "# Data Structure\nsome data\n# Current Plan\nsome plan\n# Example\nsome example\n# Instruction\nsome instruction"
+            },
+        ]
+        result = serializer.serialize_req(req)
+        expected_output = json.dumps(
+            [
+                {"content": "Command Editor.read executed: file_path=test.py"},
+                {
+                    "content": "# Data Structure\n\n\n# Current Plan\nsome plan\n# Example\n\n\n# Instruction\nsome instruction"
+                },
+            ]
+        )
+        assert result == expected_output
diff --git a/tests/metagpt/exp_pool/test_serializers/test_simple.py b/tests/metagpt/exp_pool/test_serializers/test_simple.py
new file mode 100644
index 000000000..05ef1ca11
--- /dev/null
+++ b/tests/metagpt/exp_pool/test_serializers/test_simple.py
@@ -0,0 +1,44 @@
+import pytest
+
+from metagpt.exp_pool.serializers.simple import SimpleSerializer
+
+
+class TestSimpleSerializer:
+    @pytest.fixture
+    def serializer(self):
+        return SimpleSerializer()
+
+    def test_serialize_req(self, serializer):
+        # Test with different types of input
+        assert serializer.serialize_req(123) == "123"
+        assert serializer.serialize_req("test") == "test"
+        assert serializer.serialize_req([1, 2, 3]) == "[1, 2, 3]"
+        assert serializer.serialize_req({"a": 1}) == "{'a': 1}"
+
+    def test_serialize_resp(self, serializer):
+        # Test with different types of input
+        assert serializer.serialize_resp(456) == "456"
+        assert serializer.serialize_resp("response") == "response"
+        assert serializer.serialize_resp([4, 5, 6]) == "[4, 5, 6]"
+        assert serializer.serialize_resp({"b": 2}) == "{'b': 2}"
+
+    def test_deserialize_resp(self, serializer):
+        # Test with different types of input
+        assert serializer.deserialize_resp("789") == "789"
+        assert serializer.deserialize_resp("test_response") == "test_response"
+        assert serializer.deserialize_resp("[7, 8, 9]") == "[7, 8, 9]"
+        assert serializer.deserialize_resp("{'c': 3}") == "{'c': 3}"
+
+    def test_roundtrip(self, serializer):
+        # Test serialization and deserialization roundtrip
+        original = "test_roundtrip"
+        serialized = serializer.serialize_resp(original)
+        deserialized = serializer.deserialize_resp(serialized)
+        assert deserialized == original
+
+    @pytest.mark.parametrize("input_value", [123, "test", [1, 2, 3], {"a": 1}, None])
+    def test_serialize_req_types(self, serializer, input_value):
+        # Test serialize_req with various input types
+        result = serializer.serialize_req(input_value)
+        assert isinstance(result, str)
+        assert result == str(input_value)

From bf21bbf12e43ba9188012ba1317e0491ab2c8cc9 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Fri, 12 Jul 2024 19:30:33 +0800
Subject: [PATCH 42/51] determine the content to be saved to the experience
 pool using cmd_prompt_exp_part.

---
 .../exp_pool/init_exp_pool/engineer_exps.py   |  7 +-
 .../init_exp_pool/team_leader_exps.py         | 22 +++---
 .../exp_pool/context_builders/action_node.py  | 12 +--
 metagpt/exp_pool/context_builders/base.py     |  7 +-
 .../exp_pool/context_builders/role_zero.py    | 26 ++++---
 metagpt/exp_pool/context_builders/simple.py   |  6 +-
 metagpt/exp_pool/decorator.py                 |  2 +-
 metagpt/exp_pool/serializers/role_zero.py     | 25 ++-----
 metagpt/prompts/di/role_zero.py               | 24 +++---
 metagpt/roles/di/role_zero.py                 | 22 ++++--
 .../test_rolezero_context_builder.py          | 18 +++--
 .../test_simple_context_builder.py            |  8 +-
 .../test_serializers/test_role_zero.py        | 73 ++++++-------------
 13 files changed, 113 insertions(+), 139 deletions(-)

diff --git a/examples/exp_pool/init_exp_pool/engineer_exps.py b/examples/exp_pool/init_exp_pool/engineer_exps.py
index e494fdaa3..022a0c829 100644
--- a/examples/exp_pool/init_exp_pool/engineer_exps.py
+++ b/examples/exp_pool/init_exp_pool/engineer_exps.py
@@ -3,20 +3,19 @@ import json
 OBJ1 = [
     {
         "role": "user",
-        "content": '\n# Data Structure\n\n\n# Current Plan\n{\'goal\': "Please write a 1024 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at \'/tmp/system_design.json\' and the project schedule at \'/tmp/project_schedule.json\' for detailed information.", \'tasks\': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON\'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named "src". If you know the project path, then write in a "src" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW "Data structures and interfaces". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON\'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON\'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON\'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        "command_name": str,\n        "args": {"arg_name": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n',
+        "content": "\n# Current Plan\n{'goal': \"Please write a 1048 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at '/tmp/system_design.json' and the project schedule at '/tmp/project_schedule.json' for detailed information.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named \"src\". If you know the project path, then write in a \"src\" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW \"Data structures and interfaces\". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n",
     }
 ]
 OBJ2 = [
     {
         "role": "user",
-        "content": "Command Editor.read executed: file_path='/tmp/system_design.json' block_content=\"...\"\n\nCommand Editor.read executed: file_path='/tmp/project_schedule.json' block_content=\"...\"",
+        "content": 'Command Editor.read executed: file_path=\'/tmp/system_design.json\' block_content=\'001|{"Implementation approach":"We will implement the 2048 game using plain JavaScript and HTML, ensuring no frameworks are used. The game logic will handle tile movements, merging, and game state updates. The UI will be simple and clean, with a responsive design to fit different screen sizes. We will use CSS for styling and ensure the game is playable with keyboard arrow keys. The game will display the current score, have a restart button, and show a game over message when no more moves are possible.","File list":["index.html","style.css","script.js"],"Data structures and interfaces":"\\\\nclassDiagram\\\\n    class Game {\\\\n        -grid: int[][]\\\\n        -score: int\\\\n        +init(): void\\\\n        +move(direction: str): void\\\\n        +merge(direction: str): void\\\\n        +isGameOver(): bool\\\\n        +restart(): void\\\\n    }\\\\n    class UI {\\\\n        -game: Game\\\\n        +init(): void\\\\n        +update(): void\\\\n        +showGameOver(): void\\\\n        +bindEvents(): void\\\\n    }\\\\n    Game --> UI\\\\n","Program call flow":"\\\\nsequenceDiagram\\\\n    participant U as UI\\\\n    participant G as Game\\\\n    U->>G: init()\\\\n    G-->>U: return\\\\n    U->>U: bindEvents()\\\\n    U->>G: move(direction)\\\\n    G->>G: merge(direction)\\\\n    G->>U: update()\\\\n    U->>U: update()\\\\n    U->>G: isGameOver()\\\\n    G-->>U: return bool\\\\n    alt Game Over\\\\n        U->>U: showGameOver()\\\\n    end\\\\n    U->>G: restart()\\\\n    G-->>U: return\\\\n","Anything UNCLEAR":"Clarify if there are any specific design preferences or additional features required beyond the basic 2048 game functionality."}\\n\'\n\nCommand Editor.read executed: file_path=\'/tmp/project_schedule.json\' block_content=\'001|{"Required packages":["No third-party dependencies required"],"Required Other language third-party packages":["No third-party dependencies required"],"Logic Analysis":[["script.js","Contains Game and UI classes, and their methods: init, move, merge, isGameOver, restart, update, showGameOver, bindEvents"],["index.html","Contains the HTML structure for the game UI"],["style.css","Contains the CSS styles for the game UI"]],"Task list":["index.html","style.css","script.js"],"Full API spec":"","Shared Knowledge":"The `script.js` file will contain the core game logic and UI handling. The `index.html` file will provide the structure for the game, and `style.css` will handle the styling.","Anything UNCLEAR":"Clarify if there are any specific design preferences or additional features required beyond the basic 2048 game functionality."}\\n\'',
     },
     {
         "role": "user",
-        "content": '\n# Data Structure\n\n\n# Current Plan\n{\'goal\': "Please write a 1024 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at \'/tmp/system_design.json\' and the project schedule at \'/tmp/project_schedule.json\' for detailed information.", \'tasks\': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON\'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named "src". If you know the project path, then write in a "src" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW "Data structures and interfaces". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON\'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON\'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON\'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        "command_name": str,\n        "args": {"arg_name": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n',
+        "content": "\n# Current Plan\n{'goal': \"Please write a 1048 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at '/tmp/system_design.json' and the project schedule at '/tmp/project_schedule.json' for detailed information.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named \"src\". If you know the project path, then write in a \"src\" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW \"Data structures and interfaces\". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n",
     },
 ]
-
 REQ1 = json.dumps(OBJ1)
 REQ2 = json.dumps(OBJ2)
 
diff --git a/examples/exp_pool/init_exp_pool/team_leader_exps.py b/examples/exp_pool/init_exp_pool/team_leader_exps.py
index da0bc3d81..347faac45 100644
--- a/examples/exp_pool/init_exp_pool/team_leader_exps.py
+++ b/examples/exp_pool/init_exp_pool/team_leader_exps.py
@@ -3,19 +3,19 @@ import json
 OBJ1 = [
     {
         "role": "user",
-        "content": "\n# Data Structure\n\n\n# Current Plan\n{'goal': \"from  to {'<all>'}: Create a cli snake game using Python.\", 'tasks': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        \"command_name\": str,\n        \"args\": {\"arg_name\": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n",
+        "content": "\n# Current Plan\n{'goal': \"from  to {'<all>'}: Write a 1024 game using JavaScript and HTML code without using any frameworks, user can play with keyboard.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n",
     }
 ]
 OBJ2 = [
     {
         "role": "user",
-        "content": "\n# Data Structure\n\n\n# Current Plan\n{'goal': \"from  to {'<all>'}: Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.\", 'tasks': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        \"command_name\": str,\n        \"args\": {\"arg_name\": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n",
+        "content": "\n# Current Plan\n{'goal': \"from  to {'<all>'}: Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n",
     }
 ]
 OBJ4 = [
     {
         "role": "user",
-        "content": "\n# Data Structure\n\n\n# Current Plan\n{'goal': \"from  to {'<all>'}: how does the project go?\", 'tasks': []}\n\n# Current Task\n\n\n# Example\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n\nPay close attention to the Example provided, you can reuse the example for your current situation if it fits.\nYou may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.\nIf you finish current task, you will automatically take the next task in the existing plan, use Plan.finish_task, DON'T append a new task.\n\n# Your commands in a json array, in the following output format. If there is nothing to do, use the pass or end command:\nSome text indicating your thoughts, such as how you should update the plan status, respond to inquiry, or seek for help. Then a json array of commands. You must output ONE and ONLY ONE json array. DON'T output multiple json arrays with thoughts between them.\n```json\n[\n    {\n        \"command_name\": str,\n        \"args\": {\"arg_name\": arg_value, ...}\n    },\n    ...\n]\n```\nNotice: your output JSON data section must start with **```json [**\n",
+        "content": "\n# Current Plan\n{'goal': \"from  to {'<all>'}: how does the project go?\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n",
     }
 ]
 
@@ -88,7 +88,7 @@ Explanation: The requirement is about software development. Assign each tasks to
     {
         "command_name": "RoleZero.reply_to_human",
         "args": {
-            "content": "I have assigned the tasks to the team members. Alice will create the PRD, Bob will design the software architecture, Eve will break down the architecture into tasks, Alex will implement the core game logic, and Edward will write comprehensive tests. The team will work on the project accordingly",
+            "content": "I have assigned the tasks to the team members. Alice will create the PRD, Bob will design the software architecture, Eve will break down the architecture into tasks, Alex will implement the core game logic, and Edward will write comprehensive tests. The team will work on the project accordingly"
         }
     },
     {
@@ -120,7 +120,7 @@ Explanation: DON'T decompose requirement if it is a DATA-RELATED task, assign a
     {
         "command_name": "RoleZero.reply_to_human",
         "args": {
-            "content": "I have assigned the task to David. He will break down the task further by himself and starts solving it.",
+            "content": "I have assigned the task to David. He will break down the task further by himself and starts solving it."
         }
     },
     {
@@ -139,15 +139,15 @@ Explanation: You received a message from Alice, the Product Manager, that she ha
     },
     {
         "command_name": "TeamLeader.publish_message",
-            "args": {
-                "content": "Please design the software architecture for the snake game based on the PRD created by Alice. The PRD is at 'docs/prd/20240424153821.json'. Include the choice of programming language, libraries, and data flow, etc.",
-                "send_to": "Bob"
-            }
+        "args": {
+            "content": "Please design the software architecture for the snake game based on the PRD created by Alice. The PRD is at 'docs/prd/20240424153821.json'. Include the choice of programming language, libraries, and data flow, etc.",
+            "send_to": "Bob"
+        }
     },
     {
         "command_name": "RoleZero.reply_to_human",
         "args": {
-            "content": "Alice has completed the PRD. I have marked her task as finished and sent the PRD to Bob. Bob will work on the software architecture.",
+            "content": "Alice has completed the PRD. I have marked her task as finished and sent the PRD to Bob. Bob will work on the software architecture."
         }
     },
     {
@@ -163,7 +163,7 @@ Explanation: The user is asking for a general update on the project status. Give
     {
         "command_name": "RoleZero.reply_to_human",
         "args": {
-            "content": "The team is currently working on ... We have completed ...",
+            "content": "The team is currently working on ... We have completed ..."
         }
     },
     {
diff --git a/metagpt/exp_pool/context_builders/action_node.py b/metagpt/exp_pool/context_builders/action_node.py
index ade157822..a3362875c 100644
--- a/metagpt/exp_pool/context_builders/action_node.py
+++ b/metagpt/exp_pool/context_builders/action_node.py
@@ -1,5 +1,6 @@
 """Action Node context builder."""
 
+from typing import Any
 
 from metagpt.exp_pool.context_builders.base import BaseContextBuilder
 
@@ -17,17 +18,12 @@ Consider **Experiences** to generate a better answer.
 
 
 class ActionNodeContextBuilder(BaseContextBuilder):
-    async def build(self, **kwargs) -> str:
+    async def build(self, req: Any) -> str:
         """Builds the action node context string.
 
-        Args:
-            **kwargs: Arbitrary keyword arguments, expecting 'req' as a key.
-
-        Returns:
-            str: The formatted context string using the request and formatted experiences.
-                 If no experiences are available, returns the request as is.
+        If there are no experiences, returns the original `req`;
+        otherwise returns context with `req` and formatted experiences.
         """
-        req = kwargs.get("req", "")
         exps = self.format_exps()
 
         return ACTION_NODE_CONTEXT_TEMPLATE.format(req=req, exps=exps) if exps else req
diff --git a/metagpt/exp_pool/context_builders/base.py b/metagpt/exp_pool/context_builders/base.py
index d1133c2da..f937f5c7b 100644
--- a/metagpt/exp_pool/context_builders/base.py
+++ b/metagpt/exp_pool/context_builders/base.py
@@ -16,8 +16,11 @@ class BaseContextBuilder(BaseModel, ABC):
     exps: list[Experience] = []
 
     @abstractmethod
-    async def build(self, **kwargs) -> Any:
-        """Build context from parameters."""
+    async def build(self, req: Any) -> Any:
+        """Build context from req.
+
+        Do not modify `req`. If modification is necessary, use copy.deepcopy to create a copy first.
+        """
 
     def format_exps(self) -> str:
         """Format experiences into a numbered list of strings.
diff --git a/metagpt/exp_pool/context_builders/role_zero.py b/metagpt/exp_pool/context_builders/role_zero.py
index 6407314ac..2ee469661 100644
--- a/metagpt/exp_pool/context_builders/role_zero.py
+++ b/metagpt/exp_pool/context_builders/role_zero.py
@@ -1,34 +1,36 @@
 """RoleZero context builder."""
 
+import copy
 import re
+from typing import Any
 
 from metagpt.exp_pool.context_builders.base import BaseContextBuilder
 
 
 class RoleZeroContextBuilder(BaseContextBuilder):
-    async def build(self, **kwargs) -> list[dict]:
-        """Builds the context by updating the req with formatted experiences.
+    async def build(self, req: Any) -> list[dict]:
+        """Builds the role zero context string.
 
-        Args:
-            **kwargs: Arbitrary keyword arguments, expecting 'req' as a key.
-
-        Returns:
-            list[dict]: The updated request with formatted experiences or the original request if no experiences are available.
+        Note:
+            1. The expected format for `req`, e.g., [{...}, {"role": "user", "content": "context"}, {"role": "user", "content": "context exp part"}].
+            2. Returns the original `req` if it is empty, incorrectly formatted or there are no experiences.
+            3. Creates a copy of req and replaces the example content in the copied req with actual experiences.
         """
-        req = kwargs.get("req", [])
-        if not req:
+        if not req or len(req) < 2:
             return req
 
         exps = self.format_exps()
         if not exps:
             return req
 
-        req[-1]["content"] = self.replace_example_content(req[-1].get("content", ""), exps)
+        req_copy = copy.deepcopy(req)
 
-        return req
+        req_copy[-2]["content"] = self.replace_example_content(req_copy[-2].get("content", ""), exps)
+
+        return req_copy
 
     def replace_example_content(self, text: str, new_example_content: str) -> str:
-        return self.replace_content_between_markers(text, "# Example", "# Instruction", new_example_content)
+        return self.replace_content_between_markers(text, "# Example", "# Available Commands", new_example_content)
 
     @staticmethod
     def replace_content_between_markers(text: str, start_marker: str, end_marker: str, new_content: str) -> str:
diff --git a/metagpt/exp_pool/context_builders/simple.py b/metagpt/exp_pool/context_builders/simple.py
index 565855664..d7b8d0be9 100644
--- a/metagpt/exp_pool/context_builders/simple.py
+++ b/metagpt/exp_pool/context_builders/simple.py
@@ -1,6 +1,8 @@
 """Simple context builder."""
 
 
+from typing import Any
+
 from metagpt.exp_pool.context_builders.base import BaseContextBuilder
 
 SIMPLE_CONTEXT_TEMPLATE = """
@@ -20,5 +22,5 @@ Consider **Experiences** to generate a better answer.
 
 
 class SimpleContextBuilder(BaseContextBuilder):
-    async def build(self, **kwargs) -> str:
-        return SIMPLE_CONTEXT_TEMPLATE.format(req=kwargs.get("req", ""), exps=self.format_exps())
+    async def build(self, req: Any) -> str:
+        return SIMPLE_CONTEXT_TEMPLATE.format(req=req, exps=self.format_exps())
diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 0a9a83818..566127f59 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -200,7 +200,7 @@ class ExpCacheHandler(BaseModel):
     async def _build_context(self) -> str:
         self.context_builder.exps = self._exps
 
-        return await self.context_builder.build(**self.kwargs)
+        return await self.context_builder.build(self.kwargs["req"])
 
     async def _execute_function(self):
         self.kwargs["req"] = await self._build_context()
diff --git a/metagpt/exp_pool/serializers/role_zero.py b/metagpt/exp_pool/serializers/role_zero.py
index 82a32e36b..f5363b1ff 100644
--- a/metagpt/exp_pool/serializers/role_zero.py
+++ b/metagpt/exp_pool/serializers/role_zero.py
@@ -3,7 +3,6 @@
 import copy
 import json
 
-from metagpt.exp_pool.context_builders import RoleZeroContextBuilder
 from metagpt.exp_pool.serializers.simple import SimpleSerializer
 
 
@@ -11,14 +10,15 @@ class RoleZeroSerializer(SimpleSerializer):
     def serialize_req(self, req: list[dict]) -> str:
         """Serialize the request for database storage, ensuring it is a string.
 
-        This function does not modify `req`; it only extracts the necessary content from `req` because `req` may be very lengthy and could cause embedding errors.
+        Only extracts the necessary content from `req` because `req` may be very lengthy and could cause embedding errors.
 
         Args:
             req (list[dict]): The request to be serialized. Example:
                 [
                     {"role": "user", "content": "..."},
                     {"role": "assistant", "content": "..."},
-                    {"role": "user", "content": "..."},
+                    {"role": "user", "content": "context"},
+                    {"role": "user", "content": "context exp part"},
                 ]
 
         Returns:
@@ -28,12 +28,12 @@ class RoleZeroSerializer(SimpleSerializer):
             return ""
 
         filtered_req = self._filter_req(req)
-        self._clean_last_entry_content(filtered_req)
+        filtered_req.append(req[-1])
 
         return json.dumps(filtered_req)
 
     def _filter_req(self, req: list[dict]) -> list[dict]:
-        """Filter the request to include only necessary items and the last entry.
+        """Filter the `req` to include only necessary items.
 
         Args:
             req (list[dict]): The original request.
@@ -45,20 +45,5 @@ class RoleZeroSerializer(SimpleSerializer):
         filtered_req = [
             copy.deepcopy(item) for item in req if "Command Editor.read executed: file_path" in item["content"]
         ]
-        filtered_req.append(copy.deepcopy(req[-1]))
 
         return filtered_req
-
-    def _clean_last_entry_content(self, req: list[dict]):
-        """Modifies the content of the last element in the request to remove unnecessary sections, making the request more concise."""
-
-        last_content = req[-1]["content"]
-
-        last_content = RoleZeroContextBuilder.replace_content_between_markers(
-            last_content, "# Data Structure", "# Current Plan", ""
-        )
-        last_content = RoleZeroContextBuilder.replace_content_between_markers(
-            last_content, "# Example", "# Instruction", ""
-        )
-
-        req[-1]["content"] = last_content
diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index 8f4a8804e..436ad7cdd 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -8,7 +8,16 @@ Note:
 2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.
 3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.
 """
+CMD_PROMPT_EXP_PART = """
+# Current Plan
+{plan_status}
 
+# Current Task
+{current_task}
+
+# Instruction
+{instruction}
+"""
 CMD_PROMPT = """
 # Data Structure
 class Task(BaseModel):
@@ -18,21 +27,14 @@ class Task(BaseModel):
     task_type: str = ""
     assignee: str = ""
 
+# Example
+{example}
+
 # Available Commands
 {available_commands}
 Special Command: Use {{"command_name": "end"}} to do nothing or indicate completion of all requirements and the end of actions.
 
-# Current Plan
-{plan_status}
-
-# Current Task
-{current_task}
-
-# Example
-{example}
-
-# Instruction
-{instruction}
+{cmd_prompt_exp_part}
 
 Pay close attention to the Example provided, you can reuse the example for your current situation if it fits.
 You may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 59c58861f..5b002d994 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -16,6 +16,7 @@ from metagpt.exp_pool.serializers import RoleZeroSerializer
 from metagpt.logs import logger
 from metagpt.prompts.di.role_zero import (
     CMD_PROMPT,
+    CMD_PROMPT_EXP_PART,
     JSON_REPAIR_PROMPT,
     ROLE_INSTRUCTION,
 )
@@ -144,13 +145,14 @@ class RoleZero(Role):
         tool_info = json.dumps({tool.name: tool.schemas for tool in tools})
 
         ### Make Decision Dynamically ###
-        prompt = self.cmd_prompt.format(
+        cmd_prompt_exp_part = CMD_PROMPT_EXP_PART.format(
             plan_status=plan_status,
             current_task=current_task,
-            example=example,
-            available_commands=tool_info,
             instruction=self.instruction.strip(),
         )
+        prompt = self.cmd_prompt.format(
+            example=example, available_commands=tool_info, cmd_prompt_exp_part=cmd_prompt_exp_part
+        )
         memory = self.rc.memory.get(self.memory_k)
         if not self.browser.is_empty_page:
             pattern = re.compile(r"Command Browser\.(\w+) executed")
@@ -158,16 +160,24 @@ class RoleZero(Role):
                 if pattern.match(msg.content):
                     memory.insert(index, UserMessage(cause_by="browser", content=await self.browser.view()))
                     break
-        context = self.llm.format_msg(memory + [UserMessage(content=prompt)])
-        # print(*context, sep="\n" + "*" * 5 + "\n")
+        req = self.llm.format_msg(memory + [UserMessage(content=prompt), UserMessage(content=cmd_prompt_exp_part)])
         async with ThoughtReporter(enable_llm_stream=True):
-            self.command_rsp = await self.llm_cached_aask(req=context, system_msgs=self.system_msg)
+            self.command_rsp = await self.llm_cached_aask(req=req, system_msgs=self.system_msg)
         self.rc.memory.add(AIMessage(content=self.command_rsp))
 
         return True
 
     @exp_cache(context_builder=RoleZeroContextBuilder(), serializer=RoleZeroSerializer())
     async def llm_cached_aask(self, *, req: list[dict], system_msgs: list[str]) -> str:
+        """Use `exp_cache` to automatically manage experiences.
+
+        The `RoleZeroContextBuilder` attempts to add experiences to `req`.
+        The `RoleZeroSerializer` extracts essential parts of `req` for the experience pool, trimming lengthy entries to retain only necessary parts.
+        """
+        # Remove the "cmd_prompt_exp_part", it is only used within the exp_cache decorator.
+        if req:
+            req.pop()
+
         return await self.llm.aask(req, system_msgs=system_msgs)
 
     async def _act(self) -> Message:
diff --git a/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py b/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
index 611d68211..a95566ed1 100644
--- a/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
+++ b/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
@@ -25,27 +25,31 @@ class TestRoleZeroContextBuilder:
     async def test_build_with_experiences(self, context_builder, mocker):
         mocker.patch.object(BaseContextBuilder, "format_exps", return_value="Formatted experiences")
         mocker.patch.object(RoleZeroContextBuilder, "replace_example_content", return_value="Updated content")
-        req = [{"content": "Original content"}]
+        req = [{"content": "Original content 1"}, {"content": "Original content exp part"}]
         result = await context_builder.build(req=req)
-        assert result == [{"content": "Updated content"}]
+        assert result == [{"content": "Updated content"}, {"content": "Original content exp part"}]
 
     def test_replace_example_content(self, context_builder, mocker):
         mocker.patch.object(RoleZeroContextBuilder, "replace_content_between_markers", return_value="Replaced content")
         result = context_builder.replace_example_content("Original text", "New example content")
         assert result == "Replaced content"
         context_builder.replace_content_between_markers.assert_called_once_with(
-            "Original text", "# Example", "# Instruction", "New example content"
+            "Original text", "# Example", "# Available Commands", "New example content"
         )
 
     def test_replace_content_between_markers(self):
-        text = "Start\n# Example\nOld content\n# Instruction\nEnd"
+        text = "Start\n# Example\nOld content\n# Available Commands\nEnd"
         new_content = "New content"
-        result = RoleZeroContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
-        expected = "Start\n# Example\nNew content\n\n# Instruction\nEnd"
+        result = RoleZeroContextBuilder.replace_content_between_markers(
+            text, "# Example", "# Available Commands", new_content
+        )
+        expected = "Start\n# Example\nNew content\n\n# Available Commands\nEnd"
         assert result == expected
 
     def test_replace_content_between_markers_no_match(self):
         text = "Start\nNo markers\nEnd"
         new_content = "New content"
-        result = RoleZeroContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
+        result = RoleZeroContextBuilder.replace_content_between_markers(
+            text, "# Example", "# Available Commands", new_content
+        )
         assert result == text
diff --git a/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py b/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py
index b6d0f642e..cf1a42f27 100644
--- a/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py
+++ b/tests/metagpt/exp_pool/test_context_builders/test_simple_context_builder.py
@@ -13,7 +13,7 @@ class TestSimpleContextBuilder:
         return SimpleContextBuilder()
 
     @pytest.mark.asyncio
-    async def test_build_with_experiences(self, context_builder, mocker):
+    async def test_build_with_experiences(self, mocker, context_builder: SimpleContextBuilder):
         # Mock the format_exps method
         mock_exps = "Mocked experiences"
         mocker.patch.object(BaseContextBuilder, "format_exps", return_value=mock_exps)
@@ -25,7 +25,7 @@ class TestSimpleContextBuilder:
         assert result == expected
 
     @pytest.mark.asyncio
-    async def test_build_without_experiences(self, context_builder, mocker):
+    async def test_build_without_experiences(self, mocker, context_builder: SimpleContextBuilder):
         # Mock the format_exps method to return an empty string
         mocker.patch.object(BaseContextBuilder, "format_exps", return_value="")
 
@@ -36,12 +36,12 @@ class TestSimpleContextBuilder:
         assert result == expected
 
     @pytest.mark.asyncio
-    async def test_build_without_req(self, context_builder, mocker):
+    async def test_build_without_req(self, mocker, context_builder: SimpleContextBuilder):
         # Mock the format_exps method
         mock_exps = "Mocked experiences"
         mocker.patch.object(BaseContextBuilder, "format_exps", return_value=mock_exps)
 
-        result = await context_builder.build()
+        result = await context_builder.build(req="")
 
         expected = SIMPLE_CONTEXT_TEMPLATE.format(req="", exps=mock_exps)
         assert result == expected
diff --git a/tests/metagpt/exp_pool/test_serializers/test_role_zero.py b/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
index 4c1f3daf3..d4525d535 100644
--- a/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
+++ b/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
@@ -7,71 +7,42 @@ from metagpt.exp_pool.serializers import RoleZeroSerializer
 
 class TestRoleZeroSerializer:
     @pytest.fixture
-    def serializer(self):
+    def serializer(self) -> RoleZeroSerializer:
         return RoleZeroSerializer()
 
+    @pytest.fixture
+    def last_item(self) -> dict:
+        return {
+            "role": "user",
+            "content": "# Current Plan\nsome plan\n# Current Plan\nsome plan\n# Instruction\nsome instruction",
+        }
+
+    @pytest.fixture
+    def sample_req(self):
+        return [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]
+
     def test_serialize_req_empty_input(self, serializer: RoleZeroSerializer):
         assert serializer.serialize_req([]) == ""
 
-    def test_serialize_req_with_content(self, serializer: RoleZeroSerializer):
+    def test_serialize_req_with_content(self, serializer: RoleZeroSerializer, last_item: dict):
         req = [
-            {"content": "Command Editor.read executed: file_path=test.py"},
-            {"content": "Some other content"},
-            {
-                "content": "# Data Structure\nsome data\n# Current Plan\nsome plan\n# Example\nsome example\n# Instruction\nsome instruction"
-            },
+            {"role": "user", "content": "Command Editor.read executed: file_path=test.py"},
+            {"role": "assistant", "content": "Some other content"},
+            last_item,
         ]
         expected_output = json.dumps(
-            [
-                {"content": "Command Editor.read executed: file_path=test.py"},
-                {
-                    "content": "# Data Structure\n\n\n# Current Plan\nsome plan\n# Example\n\n\n# Instruction\nsome instruction"
-                },
-            ]
+            [{"role": "user", "content": "Command Editor.read executed: file_path=test.py"}, last_item]
         )
         assert serializer.serialize_req(req) == expected_output
 
     def test_filter_req(self, serializer: RoleZeroSerializer):
         req = [
-            {"content": "Command Editor.read executed: file_path=test1.py"},
-            {"content": "Some other content"},
-            {"content": "Command Editor.read executed: file_path=test2.py"},
-            {"content": "Final content"},
+            {"role": "user", "content": "Command Editor.read executed: file_path=test1.py"},
+            {"role": "assistant", "content": "Some other content"},
+            {"role": "user", "content": "Command Editor.read executed: file_path=test2.py"},
+            {"role": "assistant", "content": "Final content"},
         ]
         filtered_req = serializer._filter_req(req)
-        assert len(filtered_req) == 3
+        assert len(filtered_req) == 2
         assert filtered_req[0]["content"] == "Command Editor.read executed: file_path=test1.py"
         assert filtered_req[1]["content"] == "Command Editor.read executed: file_path=test2.py"
-        assert filtered_req[2]["content"] == "Final content"
-
-    def test_clean_last_entry_content(self, serializer: RoleZeroSerializer):
-        req = [
-            {"content": "Some content"},
-            {
-                "content": "# Data Structure\nsome data\n# Current Plan\nsome plan\n# Example\nsome example\n# Instruction\nsome instruction"
-            },
-        ]
-        serializer._clean_last_entry_content(req)
-        expected_content = (
-            "# Data Structure\n\n\n# Current Plan\nsome plan\n# Example\n\n\n# Instruction\nsome instruction"
-        )
-        assert req[-1]["content"] == expected_content
-
-    def test_integration(self, serializer: RoleZeroSerializer):
-        req = [
-            {"content": "Command Editor.read executed: file_path=test.py"},
-            {"content": "Some other content"},
-            {
-                "content": "# Data Structure\nsome data\n# Current Plan\nsome plan\n# Example\nsome example\n# Instruction\nsome instruction"
-            },
-        ]
-        result = serializer.serialize_req(req)
-        expected_output = json.dumps(
-            [
-                {"content": "Command Editor.read executed: file_path=test.py"},
-                {
-                    "content": "# Data Structure\n\n\n# Current Plan\nsome plan\n# Example\n\n\n# Instruction\nsome instruction"
-                },
-            ]
-        )
-        assert result == expected_output

From b20316d6cd2b87a2888fbf71c2594cb3520f36e7 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 15 Jul 2024 16:00:05 +0800
Subject: [PATCH 43/51] format code

---
 metagpt/exp_pool/context_builders/action_node.py | 1 +
 metagpt/exp_pool/context_builders/role_zero.py   | 1 +
 metagpt/exp_pool/manager.py                      | 6 +++++-
 metagpt/exp_pool/scorers/simple.py               | 1 +
 metagpt/exp_pool/serializers/role_zero.py        | 1 +
 5 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/metagpt/exp_pool/context_builders/action_node.py b/metagpt/exp_pool/context_builders/action_node.py
index a3362875c..891b898be 100644
--- a/metagpt/exp_pool/context_builders/action_node.py
+++ b/metagpt/exp_pool/context_builders/action_node.py
@@ -24,6 +24,7 @@ class ActionNodeContextBuilder(BaseContextBuilder):
         If there are no experiences, returns the original `req`;
         otherwise returns context with `req` and formatted experiences.
         """
+
         exps = self.format_exps()
 
         return ACTION_NODE_CONTEXT_TEMPLATE.format(req=req, exps=exps) if exps else req
diff --git a/metagpt/exp_pool/context_builders/role_zero.py b/metagpt/exp_pool/context_builders/role_zero.py
index 2ee469661..aa5524ab4 100644
--- a/metagpt/exp_pool/context_builders/role_zero.py
+++ b/metagpt/exp_pool/context_builders/role_zero.py
@@ -16,6 +16,7 @@ class RoleZeroContextBuilder(BaseContextBuilder):
             2. Returns the original `req` if it is empty, incorrectly formatted or there are no experiences.
             3. Creates a copy of req and replaces the example content in the copied req with actual experiences.
         """
+
         if not req or len(req) < 2:
             return req
 
diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index ba1a8bcf0..d6922ff00 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -39,7 +39,7 @@ class ExperienceManager(BaseModel):
                     similarity_top_k=DEFAULT_SIMILARITY_TOP_K,
                 )
             ]
-            ranker_configs = [LLMRankerConfig()]
+            ranker_configs = [LLMRankerConfig(top_n=DEFAULT_SIMILARITY_TOP_K)]
 
             self.storage = SimpleEngine.from_objs(retriever_configs=retriever_configs, ranker_configs=ranker_configs)
 
@@ -57,6 +57,7 @@ class ExperienceManager(BaseModel):
         Args:
             exp (Experience): The experience to add.
         """
+
         if not self.config.exp_pool.enable_write:
             return
 
@@ -74,6 +75,7 @@ class ExperienceManager(BaseModel):
         Returns:
             list[Experience]: A list of experiences that match the args.
         """
+
         if not self.config.exp_pool.enable_read:
             return []
 
@@ -90,6 +92,8 @@ class ExperienceManager(BaseModel):
         return exps
 
     def get_exps_count(self) -> int:
+        """Get the total number of experiences."""
+
         return self.vector_store._collection.count()
 
 
diff --git a/metagpt/exp_pool/scorers/simple.py b/metagpt/exp_pool/scorers/simple.py
index fd7b6537b..4b060aac4 100644
--- a/metagpt/exp_pool/scorers/simple.py
+++ b/metagpt/exp_pool/scorers/simple.py
@@ -57,6 +57,7 @@ class SimpleScorer(BaseScorer):
         Returns:
             Score: An object containing the score (1-10) and the reasoning.
         """
+
         prompt = SIMPLE_SCORER_TEMPLATE.format(req=req, resp=resp)
         resp = await self.llm.aask(prompt)
         resp_json = json.loads(CodeParser.parse_code(resp, lang="json"))
diff --git a/metagpt/exp_pool/serializers/role_zero.py b/metagpt/exp_pool/serializers/role_zero.py
index f5363b1ff..720bf5078 100644
--- a/metagpt/exp_pool/serializers/role_zero.py
+++ b/metagpt/exp_pool/serializers/role_zero.py
@@ -24,6 +24,7 @@ class RoleZeroSerializer(SimpleSerializer):
         Returns:
             str: The serialized request as a JSON string.
         """
+
         if not req:
             return ""
 

From e315f61f60fe1a3bcbcdde3df115a2314673a2dd Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 15 Jul 2024 17:10:34 +0800
Subject: [PATCH 44/51] update comment

---
 metagpt/exp_pool/context_builders/base.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/metagpt/exp_pool/context_builders/base.py b/metagpt/exp_pool/context_builders/base.py
index f937f5c7b..691d51c8c 100644
--- a/metagpt/exp_pool/context_builders/base.py
+++ b/metagpt/exp_pool/context_builders/base.py
@@ -7,7 +7,7 @@ from pydantic import BaseModel, ConfigDict
 
 from metagpt.exp_pool.schema import Experience
 
-EXP_TEMPLATE = """Given the request: {req}, We can get the response: {resp}, Which scored: {score}."""
+EXP_TEMPLATE = """Given the request: {req}, We can get the response: {resp}, which scored: {score}."""
 
 
 class BaseContextBuilder(BaseModel, ABC):
@@ -26,8 +26,8 @@ class BaseContextBuilder(BaseModel, ABC):
         """Format experiences into a numbered list of strings.
 
         Example:
-            1. Given the request: req1, We can get the response: resp1, Which scored: 8.
-            2. Given the request: req2, We can get the response: resp2, Which scored: 9.
+            1. Given the request: req1, We can get the response: resp1, which scored: 8.
+            2. Given the request: req2, We can get the response: resp2, which scored: 9.
 
         Returns:
             str: The formatted experiences as a string.

From 0c6786feaaac5fa3cd9a684877f5c68450c285d5 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Mon, 15 Jul 2024 18:22:05 +0800
Subject: [PATCH 45/51] update comment

---
 metagpt/exp_pool/serializers/role_zero.py | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/metagpt/exp_pool/serializers/role_zero.py b/metagpt/exp_pool/serializers/role_zero.py
index 720bf5078..967e53e41 100644
--- a/metagpt/exp_pool/serializers/role_zero.py
+++ b/metagpt/exp_pool/serializers/role_zero.py
@@ -43,8 +43,14 @@ class RoleZeroSerializer(SimpleSerializer):
             list[dict]: The filtered request.
         """
 
-        filtered_req = [
-            copy.deepcopy(item) for item in req if "Command Editor.read executed: file_path" in item["content"]
-        ]
+        filtered_req = [copy.deepcopy(item) for item in req if self._is_useful_content(item["content"])]
 
         return filtered_req
+
+    def _is_useful_content(self, content: str) -> bool:
+        """Currently, only the content of the file is considered, and more judgments can be added later."""
+
+        if "Command Editor.read executed: file_path" in content:
+            return True
+
+        return False

From 8b9e992b566bf2bab8b582e219200ec5e8a46a84 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 16 Jul 2024 16:39:43 +0800
Subject: [PATCH 46/51] update serializers

---
 metagpt/exp_pool/decorator.py             |  2 +-
 metagpt/exp_pool/serializers/base.py      |  4 ++--
 metagpt/exp_pool/serializers/role_zero.py |  8 +++++---
 metagpt/exp_pool/serializers/simple.py    |  4 ++--
 metagpt/prompts/di/role_zero.py           | 19 ++++++++---------
 metagpt/roles/di/role_zero.py             | 25 +++++++++++------------
 6 files changed, 30 insertions(+), 32 deletions(-)

diff --git a/metagpt/exp_pool/decorator.py b/metagpt/exp_pool/decorator.py
index 566127f59..7a2f926c5 100644
--- a/metagpt/exp_pool/decorator.py
+++ b/metagpt/exp_pool/decorator.py
@@ -121,7 +121,7 @@ class ExpCacheHandler(BaseModel):
         self.serializer = self.serializer or SimpleSerializer()
         self.tag = self.tag or self._generate_tag()
 
-        self._req = self.serializer.serialize_req(self.kwargs["req"])
+        self._req = self.serializer.serialize_req(**self.kwargs)
 
         return self
 
diff --git a/metagpt/exp_pool/serializers/base.py b/metagpt/exp_pool/serializers/base.py
index 9d00a05b2..c09488e12 100644
--- a/metagpt/exp_pool/serializers/base.py
+++ b/metagpt/exp_pool/serializers/base.py
@@ -10,10 +10,10 @@ class BaseSerializer(BaseModel, ABC):
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
     @abstractmethod
-    def serialize_req(self, req: Any) -> str:
+    def serialize_req(self, **kwargs) -> str:
         """Serializes the request for storage.
 
-        Do not modify req. If modification is necessary, use copy.deepcopy to create a copy first.
+        Do not modify kwargs. If modification is necessary, use copy.deepcopy to create a copy first.
         Note that copy.deepcopy may raise errors, such as TypeError: cannot pickle '_thread.RLock' object.
         """
 
diff --git a/metagpt/exp_pool/serializers/role_zero.py b/metagpt/exp_pool/serializers/role_zero.py
index 967e53e41..89dd73f39 100644
--- a/metagpt/exp_pool/serializers/role_zero.py
+++ b/metagpt/exp_pool/serializers/role_zero.py
@@ -7,7 +7,7 @@ from metagpt.exp_pool.serializers.simple import SimpleSerializer
 
 
 class RoleZeroSerializer(SimpleSerializer):
-    def serialize_req(self, req: list[dict]) -> str:
+    def serialize_req(self, **kwargs) -> str:
         """Serialize the request for database storage, ensuring it is a string.
 
         Only extracts the necessary content from `req` because `req` may be very lengthy and could cause embedding errors.
@@ -18,18 +18,20 @@ class RoleZeroSerializer(SimpleSerializer):
                     {"role": "user", "content": "..."},
                     {"role": "assistant", "content": "..."},
                     {"role": "user", "content": "context"},
-                    {"role": "user", "content": "context exp part"},
                 ]
 
         Returns:
             str: The serialized request as a JSON string.
         """
+        req = kwargs.get("req", [])
 
         if not req:
             return ""
 
         filtered_req = self._filter_req(req)
-        filtered_req.append(req[-1])
+
+        if state_data := kwargs.get("state_data"):
+            filtered_req.append({"role": "user", "content": state_data})
 
         return json.dumps(filtered_req)
 
diff --git a/metagpt/exp_pool/serializers/simple.py b/metagpt/exp_pool/serializers/simple.py
index 32fe29c9f..ebd06e0e0 100644
--- a/metagpt/exp_pool/serializers/simple.py
+++ b/metagpt/exp_pool/serializers/simple.py
@@ -6,10 +6,10 @@ from metagpt.exp_pool.serializers.base import BaseSerializer
 
 
 class SimpleSerializer(BaseSerializer):
-    def serialize_req(self, req: Any) -> str:
+    def serialize_req(self, **kwargs) -> str:
         """Just use `str` to convert the request object into a string."""
 
-        return str(req)
+        return str(kwargs.get("req", ""))
 
     def serialize_resp(self, resp: Any) -> str:
         """Just use `str` to convert the response object into a string."""
diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index f1de5bf1b..41b9e023e 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -8,16 +8,6 @@ Note:
 2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task by Plan.finish_current_task explicitly.
 3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.
 """
-CMD_PROMPT_EXP_PART = """
-# Current Plan
-{plan_status}
-
-# Current Task
-{current_task}
-
-# Instruction
-{instruction}
-"""
 # To ensure compatibility with hard-coded experience, do not add any other content between "# Example" and "# Available Commands".
 CMD_PROMPT = """
 # Data Structure
@@ -38,7 +28,14 @@ Special Command: Use {{"command_name": "end"}} to do nothing or indicate complet
 # Available Task Types
 {task_type_desc}
 
-{cmd_prompt_exp_part}
+# Current Plan
+{plan_status}
+
+# Current Task
+{current_task}
+
+# Instruction
+{instruction}
 
 Pay close attention to the Example provided, you can reuse the example for your current situation if it fits.
 You may use any of the available commands to create a plan or update the plan. You may output mutiple commands, they will be executed sequentially.
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 8411cf24b..ce305707e 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -16,7 +16,6 @@ from metagpt.exp_pool.serializers import RoleZeroSerializer
 from metagpt.logs import logger
 from metagpt.prompts.di.role_zero import (
     CMD_PROMPT,
-    CMD_PROMPT_EXP_PART,
     JSON_REPAIR_PROMPT,
     QUICK_THINK_PROMPT,
     ROLE_INSTRUCTION,
@@ -147,39 +146,39 @@ class RoleZero(Role):
         tool_info = json.dumps({tool.name: tool.schemas for tool in tools})
 
         ### Make Decision Dynamically ###
-        cmd_prompt_exp_part = CMD_PROMPT_EXP_PART.format(
-            plan_status=plan_status,
-            current_task=current_task,
-            instruction=self.instruction.strip(),
-        )
+        instruction = self.instruction.strip()
         prompt = self.cmd_prompt.format(
             example=example,
             available_commands=tool_info,
             task_type_desc=self.task_type_desc,
-            cmd_prompt_exp_part=cmd_prompt_exp_part,
+            plan_status=plan_status,
+            current_task=current_task,
+            instruction=instruction,
         )
         memory = self.rc.memory.get(self.memory_k)
         memory = await self.parse_browser_actions(memory)
 
-        req = self.llm.format_msg(memory + [UserMessage(content=prompt), UserMessage(content=cmd_prompt_exp_part)])
+        req = self.llm.format_msg(memory + [UserMessage(content=prompt)])
         async with ThoughtReporter(enable_llm_stream=True) as reporter:
             await reporter.async_report({"type": "react"})
-            self.command_rsp = await self.llm_cached_aask(req=req, system_msgs=self.system_msg)
+            state_data = dict(
+                plan_status=plan_status,
+                current_task=current_task,
+                instruction=instruction,
+            )
+            self.command_rsp = await self.llm_cached_aask(req=req, system_msgs=self.system_msg, state_data=state_data)
 
         self.rc.memory.add(AIMessage(content=self.command_rsp))
 
         return True
 
     @exp_cache(context_builder=RoleZeroContextBuilder(), serializer=RoleZeroSerializer())
-    async def llm_cached_aask(self, *, req: list[dict], system_msgs: list[str]) -> str:
+    async def llm_cached_aask(self, *, req: list[dict], system_msgs: list[str], **kwargs) -> str:
         """Use `exp_cache` to automatically manage experiences.
 
         The `RoleZeroContextBuilder` attempts to add experiences to `req`.
         The `RoleZeroSerializer` extracts essential parts of `req` for the experience pool, trimming lengthy entries to retain only necessary parts.
         """
-        # Remove the "cmd_prompt_exp_part", it is only used within the exp_cache decorator.
-        if req:
-            req.pop()
 
         return await self.llm.aask(req, system_msgs=system_msgs)
 

From 0b604c42b5f3af2f2fefa522ac526e51b472f8e0 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 16 Jul 2024 17:28:54 +0800
Subject: [PATCH 47/51] lazy import rag

---
 metagpt/exp_pool/manager.py | 44 +++++++++++++++++++++++++------------
 1 file changed, 30 insertions(+), 14 deletions(-)

diff --git a/metagpt/exp_pool/manager.py b/metagpt/exp_pool/manager.py
index d6922ff00..9bf289038 100644
--- a/metagpt/exp_pool/manager.py
+++ b/metagpt/exp_pool/manager.py
@@ -1,7 +1,8 @@
 """Experience Manager."""
 
-from llama_index.vector_stores.chroma import ChromaVectorStore
-from pydantic import BaseModel, ConfigDict, model_validator
+from typing import TYPE_CHECKING, Any
+
+from pydantic import BaseModel, ConfigDict
 
 from metagpt.config2 import Config, config
 from metagpt.exp_pool.schema import (
@@ -11,27 +12,37 @@ from metagpt.exp_pool.schema import (
     QueryType,
 )
 from metagpt.logs import logger
-from metagpt.rag.engines import SimpleEngine
-from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
 from metagpt.utils.exceptions import handle_exception
 
+if TYPE_CHECKING:
+    from llama_index.vector_stores.chroma import ChromaVectorStore
+
 
 class ExperienceManager(BaseModel):
     """ExperienceManager manages the lifecycle of experiences, including CRUD and optimization.
 
     Args:
         config (Config): Configuration for managing experiences.
-        storage (SimpleEngine): Engine to handle the storage and retrieval of experiences.
+        _storage (SimpleEngine): Engine to handle the storage and retrieval of experiences.
+        _vector_store (ChromaVectorStore): The actual place where vectors are stored.
     """
 
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
     config: Config = config
-    storage: SimpleEngine = None
 
-    @model_validator(mode="after")
-    def initialize(self):
-        if self.storage is None:
+    _storage: Any = None
+    _vector_store: Any = None
+
+    @property
+    def storage(self):
+        if self._storage is None:
+            try:
+                from metagpt.rag.engines import SimpleEngine
+                from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
+            except ImportError:
+                raise ImportError("To use the experience pool, you need to install the rag module.")
+
             retriever_configs = [
                 ChromaRetrieverConfig(
                     persist_path=self.config.exp_pool.persist_path,
@@ -41,14 +52,19 @@ class ExperienceManager(BaseModel):
             ]
             ranker_configs = [LLMRankerConfig(top_n=DEFAULT_SIMILARITY_TOP_K)]
 
-            self.storage = SimpleEngine.from_objs(retriever_configs=retriever_configs, ranker_configs=ranker_configs)
+            self._storage: SimpleEngine = SimpleEngine.from_objs(
+                retriever_configs=retriever_configs, ranker_configs=ranker_configs
+            )
+            logger.debug(f"exp_pool config: {self.config.exp_pool}")
 
-        logger.debug(f"exp_pool config: {self.config.exp_pool}")
-        return self
+        return self._storage
 
     @property
-    def vector_store(self) -> ChromaVectorStore:
-        return self.storage._retriever._vector_store
+    def vector_store(self):
+        if not self._vector_store:
+            self._vector_store: ChromaVectorStore = self.storage._retriever._vector_store
+
+        return self._vector_store
 
     @handle_exception
     def create_exp(self, exp: Experience):

From 5693594afe175de1ddc5702d2f29047e4004dd95 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 16 Jul 2024 19:13:10 +0800
Subject: [PATCH 48/51] add README.md

---
 examples/data/exp_pool/engineer_exps.json     |  16 ++
 examples/data/exp_pool/team_leader_exps.json  |  22 +++
 examples/exp_pool/README.md                   |  19 ++
 examples/exp_pool/decorator.py                |   6 +-
 examples/exp_pool/init_exp_pool.py            |  94 +++++++++
 .../exp_pool/init_exp_pool/engineer_exps.py   | 114 -----------
 examples/exp_pool/init_exp_pool/main.py       |  55 ------
 .../init_exp_pool/team_leader_exps.py         | 181 ------------------
 examples/exp_pool/manager.py                  |  14 +-
 examples/exp_pool/scorer.py                   |  12 ++
 tests/metagpt/exp_pool/test_manager.py        |  82 ++++----
 .../test_serializers/test_role_zero.py        |   4 +-
 .../exp_pool/test_serializers/test_simple.py  |  20 +-
 13 files changed, 238 insertions(+), 401 deletions(-)
 create mode 100644 examples/data/exp_pool/engineer_exps.json
 create mode 100644 examples/data/exp_pool/team_leader_exps.json
 create mode 100644 examples/exp_pool/README.md
 create mode 100644 examples/exp_pool/init_exp_pool.py
 delete mode 100644 examples/exp_pool/init_exp_pool/engineer_exps.py
 delete mode 100644 examples/exp_pool/init_exp_pool/main.py
 delete mode 100644 examples/exp_pool/init_exp_pool/team_leader_exps.py

diff --git a/examples/data/exp_pool/engineer_exps.json b/examples/data/exp_pool/engineer_exps.json
new file mode 100644
index 000000000..89c0bc186
--- /dev/null
+++ b/examples/data/exp_pool/engineer_exps.json
@@ -0,0 +1,16 @@
+[{
+	"req": [{
+		"role": "user",
+		"content": "\n# Current Plan\n{'goal': \"Please write a 1048 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at '/tmp/system_design.json' and the project schedule at '/tmp/project_schedule.json' for detailed information.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named \"src\". If you know the project path, then write in a \"src\" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW \"Data structures and interfaces\". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n"
+	}],
+	"resp": "\nExplanation: To begin the task of creating a 2048 game using JavaScript and HTML, I need to first read the system design and project schedule documents to understand the specific requirements and architecture outlined for the game development.\n\n```json\n[\n    {\n        \"command_name\": \"Editor.read\",\n        \"args\": {\n            \"path\": \"/tmp/system_design.json\"\n        }\n    },\n    {\n        \"command_name\": \"Editor.read\",\n        \"args\": {\n            \"path\": \"/tmp/project_schedule.json\"\n        }\n    }\n]\n```\n"
+}, {
+	"req": [{
+		"role": "user",
+		"content": "Command Editor.read executed: file_path='/tmp/system_design.json' block_content='001|{\"Implementation approach\":\"We will implement the 2048 game using plain JavaScript and HTML, ensuring no frameworks are used. The game logic will handle tile movements, merging, and game state updates. The UI will be simple and clean, with a responsive design to fit different screen sizes. We will use CSS for styling and ensure the game is playable with keyboard arrow keys. The game will display the current score, have a restart button, and show a game over message when no more moves are possible.\",\"File list\":[\"index.html\",\"style.css\",\"script.js\"],\"Data structures and interfaces\":\"\\\\nclassDiagram\\\\n    class Game {\\\\n        -grid: int[][]\\\\n        -score: int\\\\n        +init(): void\\\\n        +move(direction: str): void\\\\n        +merge(direction: str): void\\\\n        +isGameOver(): bool\\\\n        +restart(): void\\\\n    }\\\\n    class UI {\\\\n        -game: Game\\\\n        +init(): void\\\\n        +update(): void\\\\n        +showGameOver(): void\\\\n        +bindEvents(): void\\\\n    }\\\\n    Game --> UI\\\\n\",\"Program call flow\":\"\\\\nsequenceDiagram\\\\n    participant U as UI\\\\n    participant G as Game\\\\n    U->>G: init()\\\\n    G-->>U: return\\\\n    U->>U: bindEvents()\\\\n    U->>G: move(direction)\\\\n    G->>G: merge(direction)\\\\n    G->>U: update()\\\\n    U->>U: update()\\\\n    U->>G: isGameOver()\\\\n    G-->>U: return bool\\\\n    alt Game Over\\\\n        U->>U: showGameOver()\\\\n    end\\\\n    U->>G: restart()\\\\n    G-->>U: return\\\\n\",\"Anything UNCLEAR\":\"Clarify if there are any specific design preferences or additional features required beyond the basic 2048 game functionality.\"}\\n'\n\nCommand Editor.read executed: file_path='/tmp/project_schedule.json' block_content='001|{\"Required packages\":[\"No third-party dependencies required\"],\"Required Other language third-party packages\":[\"No third-party dependencies required\"],\"Logic Analysis\":[[\"script.js\",\"Contains Game and UI classes, and their methods: init, move, merge, isGameOver, restart, update, showGameOver, bindEvents\"],[\"index.html\",\"Contains the HTML structure for the game UI\"],[\"style.css\",\"Contains the CSS styles for the game UI\"]],\"Task list\":[\"index.html\",\"style.css\",\"script.js\"],\"Full API spec\":\"\",\"Shared Knowledge\":\"The `script.js` file will contain the core game logic and UI handling. The `index.html` file will provide the structure for the game, and `style.css` will handle the styling.\",\"Anything UNCLEAR\":\"Clarify if there are any specific design preferences or additional features required beyond the basic 2048 game functionality.\"}\\n'"
+	}, {
+		"role": "user",
+		"content": "\n# Current Plan\n{'goal': \"Please write a 1048 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at '/tmp/system_design.json' and the project schedule at '/tmp/project_schedule.json' for detailed information.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named \"src\". If you know the project path, then write in a \"src\" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW \"Data structures and interfaces\". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n"
+	}],
+	"resp": "\nExplanation: Based on the system design and project schedule, I will create a plan to implement the 2048 game. The plan will consist of tasks to create the necessary files (`index.html`, `style.css`, and `script.js`) and review each file to ensure they meet the design specifications.\n\nHere's the plan:\n\n1. **Task 1**: Create `index.html` - This file will contain the HTML structure necessary for the game's UI.\n2. **Task 2**: Create `style.css` - This file will define the CSS styles to make the game visually appealing and responsive.\n3. **Task 3**: Create `script.js` - This file will contain the JavaScript code for the game logic and UI interactions.\n4. **Code Review Tasks**: Review each file to ensure they meet the project requirements and adhere to the system design.\n\nLet's start by appending the first task to the plan.\n\n```json\n[\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"1\",\n            \"dependent_task_ids\": [],\n            \"instruction\": \"Create the index.html file with the basic HTML structure for the 2048 game.\",\n            \"assignee\": \"Alex\"\n        }\n    },\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"2\",\n            \"dependent_task_ids\": [\"1\"],\n            \"instruction\": \"Create the style.css file with the necessary CSS to style the 2048 game.\",\n            \"assignee\": \"Alex\"\n        }\n    },\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"3\",\n            \"dependent_task_ids\": [\"1\", \"2\"],\n            \"instruction\": \"Create the script.js file containing the JavaScript logic for the 2048 game.\",\n            \"assignee\": \"Alex\"\n        }\n    },\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"4\",\n            \"dependent_task_ids\": [\"1\"],\n            \"instruction\": \"Use ReviewAndRewriteCode to review the code in index.html to ensure it meets the design specifications.\",\n            \"assignee\": \"Alex\"\n        }\n    },\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"5\",\n            \"dependent_task_ids\": [\"2\"],\n            \"instruction\": \"Use ReviewAndRewriteCode to review the code in style.css to ensure it meets the design specifications.\",\n            \"assignee\": \"Alex\"\n        }\n    },\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"6\",\n            \"dependent_task_ids\": [\"3\"],\n            \"instruction\": \"Use ReviewAndRewriteCode to review the code in script.js to ensure it meets the design specifications.\",\n            \"assignee\": \"Alex\"\n        }\n    }\n]\n```\n"
+}]
\ No newline at end of file
diff --git a/examples/data/exp_pool/team_leader_exps.json b/examples/data/exp_pool/team_leader_exps.json
new file mode 100644
index 000000000..125f0a48d
--- /dev/null
+++ b/examples/data/exp_pool/team_leader_exps.json
@@ -0,0 +1,22 @@
+[{
+	"req": [{
+		"role": "user",
+		"content": "\n# Current Plan\n{'goal': \"from  to {'<all>'}: Write a 1024 game using JavaScript and HTML code without using any frameworks, user can play with keyboard.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n"
+	}],
+	"resp": "\nExplanation: The requirement is about software development. Assign each tasks to a different team member based on their expertise. When publishing message to Product Manager, we copy original user requirement directly to ensure no information loss.\n```json\n[\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"1\",\n            \"dependent_task_ids\": [],\n            \"instruction\": \"Create a product requirement document (PRD) outlining the features, user interface, and user experience of the CLI python snake game.\",\n            \"assignee\": \"Alice\"\n        }\n    },\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"2\",\n            \"dependent_task_ids\": [\"1\"],\n            \"instruction\": \"Design the software architecture for the CLI snake game, including the choice of programming language, libraries, and data flow.\",\n            \"assignee\": \"Bob\"\n        }\n    },\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"3\",\n            \"dependent_task_ids\": [\"2\"],\n            \"instruction\": \"Break down the architecture into manageable tasks, identify task dependencies, and prepare a detailed task list for implementation.\",\n            \"assignee\": \"Eve\"\n        }\n    },\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"4\",\n            \"dependent_task_ids\": [\"3\"],\n            \"instruction\": \"Implement the core game logic for the CLI snake game, including snake movement, food generation, and score tracking.\",\n            \"assignee\": \"Alex\"\n        }\n    },\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"5\",\n            \"dependent_task_ids\": [\"4\"],\n            \"instruction\": \"Write comprehensive tests for the game logic and user interface to ensure functionality and reliability.\",\n            \"assignee\": \"Edward\"\n        }\n    },\n    {\n        \"command_name\": \"TeamLeader.publish_message\",\n        \"args\": {\n            \"content\": \"Create a cli snake game using Python\",\n            \"send_to\": \"Alice\"\n        }\n    },\n    {\n        \"command_name\": \"RoleZero.reply_to_human\",\n        \"args\": {\n            \"content\": \"I have assigned the tasks to the team members. Alice will create the PRD, Bob will design the software architecture, Eve will break down the architecture into tasks, Alex will implement the core game logic, and Edward will write comprehensive tests. The team will work on the project accordingly\"\n        }\n    },\n    {\n        \"command_name\": \"end\"\n    }\n]\n```\n"
+}, {
+	"req": [{
+		"role": "user",
+		"content": "\n# Current Plan\n{'goal': \"from  to {'<all>'}: Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n"
+	}],
+	"resp": "\nExplanation: DON'T decompose requirement if it is a DATA-RELATED task, assign a single task directly to Data Analyst David. He will manage the decomposition and implementation.\n```json\n[\n    {\n        \"command_name\": \"Plan.append_task\",\n        \"args\": {\n            \"task_id\": \"1\",\n            \"dependent_task_ids\": [],\n            \"instruction\": \"Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.\",\n            \"assignee\": \"David\"\n        }\n    },\n    {\n        \"command_name\": \"TeamLeader.publish_message\",\n        \"args\": {\n            \"content\": \"Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.\",\n            \"send_to\": \"David\"\n        }\n    },\n    {\n        \"command_name\": \"RoleZero.reply_to_human\",\n        \"args\": {\n            \"content\": \"I have assigned the task to David. He will break down the task further by himself and starts solving it.\"\n        }\n    },\n    {\n        \"command_name\": \"end\"\n    }\n]\n```\n"
+}, {
+	"req": "Conversation History:\n[\n    ...,\n    {'role': 'assistant', 'content': 'from Alice(Product Manager) to {'Bob'}: {'docs': {'20240424153821.json': {'root_path': 'docs/prd', 'filename': '20240424153821.json', 'content': '{\"Language\":\"en_us\",\"Programming Language\":\"Python\",\"Original Requirements\":\"create a cli snake game\",\"Project Name\":\"snake_game\",\"Product Goals\":[\"Develop an intuitive and addictive snake game\",...], ...}}}}},\n]\n",
+	"resp": "\nExplanation: You received a message from Alice, the Product Manager, that she has completed the PRD, use Plan.finish_current_task to mark her task as finished and moves the plan to the next task. Based on plan status, next task is for Bob (Architect), publish a message asking him to start. The message content should contain important path info.\n```json\n[\n    {\n        \"command_name\": \"Plan.finish_current_task\",\n        \"args\": {}\n    },\n    {\n        \"command_name\": \"TeamLeader.publish_message\",\n        \"args\": {\n            \"content\": \"Please design the software architecture for the snake game based on the PRD created by Alice. The PRD is at 'docs/prd/20240424153821.json'. Include the choice of programming language, libraries, and data flow, etc.\",\n            \"send_to\": \"Bob\"\n        }\n    },\n    {\n        \"command_name\": \"RoleZero.reply_to_human\",\n        \"args\": {\n            \"content\": \"Alice has completed the PRD. I have marked her task as finished and sent the PRD to Bob. Bob will work on the software architecture.\"\n        }\n    },\n    {\n        \"command_name\": \"end\"\n    }\n]\n```\n"
+}, {
+	"req": [{
+		"role": "user",
+		"content": "\n# Current Plan\n{'goal': \"from  to {'<all>'}: how does the project go?\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n"
+	}],
+	"resp": "\nExplanation: The user is asking for a general update on the project status. Give a straight answer about the current task the team is working on and provide a summary of the completed tasks.\n```json\n[\n    {\n        \"command_name\": \"RoleZero.reply_to_human\",\n        \"args\": {\n            \"content\": \"The team is currently working on ... We have completed ...\"\n        }\n    },\n    {\n        \"command_name\": \"end\"\n    }\n]\n```\n"
+}]
\ No newline at end of file
diff --git a/examples/exp_pool/README.md b/examples/exp_pool/README.md
new file mode 100644
index 000000000..d405bfa22
--- /dev/null
+++ b/examples/exp_pool/README.md
@@ -0,0 +1,19 @@
+# Experience Pool
+
+## Prerequisites
+- Ensure the RAG module is installed: https://docs.deepwisdom.ai/main/en/guide/in_depth_guides/rag_module.html
+- Set both `enable_read` and `enable_write` to `true` in the `exp_pool` section of `config2.yaml`
+
+## Example Files
+
+### 1. decorator.py
+Showcases the implementation of the `@exp_cache` decorator.
+
+### 2. init_exp_pool.py
+Demonstrates the process of initializing the experience pool.
+
+### 3. manager.py
+Illustrates CRUD (Create, Read, Update, Delete) operations for managing experiences in the pool.
+
+### 4. scorer.py
+Outlines methods for evaluating and scoring experiences within the pool.
diff --git a/examples/exp_pool/decorator.py b/examples/exp_pool/decorator.py
index 00726a0a8..d25949e8d 100644
--- a/examples/exp_pool/decorator.py
+++ b/examples/exp_pool/decorator.py
@@ -1,4 +1,6 @@
-"""Decorator example of experience pool."""
+"""
+This script demonstrates how to automatically store experiences using @exp_cache and query the stored experiences.
+"""
 
 import asyncio
 import uuid
@@ -16,7 +18,7 @@ async def main():
     req = "Water"
 
     resp = await produce(req=req)
-    logger.info(f"The resp of `produce{req}` is: {resp}")
+    logger.info(f"The response of `produce({req})` is: {resp}")
 
     exps = await exp_manager.query_exps(req)
     logger.info(f"Find experiences: {exps}")
diff --git a/examples/exp_pool/init_exp_pool.py b/examples/exp_pool/init_exp_pool.py
new file mode 100644
index 000000000..14c415be7
--- /dev/null
+++ b/examples/exp_pool/init_exp_pool.py
@@ -0,0 +1,94 @@
+"""Init experience pool.
+
+Put some useful experiences into the experience pool.
+"""
+
+import asyncio
+import json
+from pathlib import Path
+
+from metagpt.const import EXAMPLE_DATA_PATH
+from metagpt.exp_pool import exp_manager
+from metagpt.exp_pool.schema import EntryType, Experience, Metric, Score
+from metagpt.logs import logger
+from metagpt.utils.common import aread
+
+
+async def load_file(filepath) -> list[dict]:
+    """Asynchronously loads and parses a JSON file.
+
+    Args:
+        filepath: Path to the JSON file.
+
+    Returns:
+        A list of dictionaries parsed from the JSON file.
+    """
+
+    return json.loads(await aread(filepath))
+
+
+async def add_exp(req: str, resp: str, tag: str, metric: Metric = None):
+    """Adds a new experience to the experience pool.
+
+    Args:
+        req: The request string.
+        resp: The response string.
+        tag: A tag for categorizing the experience.
+        metric: Optional metric for the experience. Defaults to a score of 10.
+
+    """
+
+    exp = Experience(
+        req=req,
+        resp=resp,
+        entry_type=EntryType.MANUAL,
+        tag=tag,
+        metric=metric or Metric(score=Score(val=10, reason="Manual")),
+    )
+
+    exp_manager.config.exp_pool.enable_write = True
+    exp_manager.create_exp(exp)
+    logger.info(f"New experience created for the request `{req[:10]}`.")
+
+
+async def add_exps(exps: list, tag: str):
+    """Adds multiple experiences to the experience pool.
+
+    Args:
+        exps: A list of experience dictionaries.
+        tag: A tag for categorizing the experiences.
+
+    """
+
+    tasks = [add_exp(req=json.dumps(exp["req"]), resp=exp["resp"], tag=tag) for exp in exps]
+    await asyncio.gather(*tasks)
+
+
+async def add_exps_from_file(tag: str, filepath: Path):
+    """Loads experiences from a file and adds them to the experience pool.
+
+    Args:
+        tag: A tag for categorizing the experiences.
+        filepath: Path to the file containing experiences.
+
+    """
+
+    exps = await load_file(filepath)
+    await add_exps(exps, tag)
+
+
+def query_exps_count():
+    """Queries and logs the total count of experiences in the pool."""
+
+    count = exp_manager.get_exps_count()
+    logger.info(f"Experiences Count: {count}")
+
+
+async def main():
+    await add_exps_from_file("TeamLeader.llm_cached_aask", EXAMPLE_DATA_PATH / "exp_pool/team_leader_exps.json"),
+    await add_exps_from_file("Engineer2.llm_cached_aask", EXAMPLE_DATA_PATH / "exp_pool/engineer_exps.json")
+    query_exps_count()
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/examples/exp_pool/init_exp_pool/engineer_exps.py b/examples/exp_pool/init_exp_pool/engineer_exps.py
deleted file mode 100644
index 022a0c829..000000000
--- a/examples/exp_pool/init_exp_pool/engineer_exps.py
+++ /dev/null
@@ -1,114 +0,0 @@
-import json
-
-OBJ1 = [
-    {
-        "role": "user",
-        "content": "\n# Current Plan\n{'goal': \"Please write a 1048 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at '/tmp/system_design.json' and the project schedule at '/tmp/project_schedule.json' for detailed information.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named \"src\". If you know the project path, then write in a \"src\" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW \"Data structures and interfaces\". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n",
-    }
-]
-OBJ2 = [
-    {
-        "role": "user",
-        "content": 'Command Editor.read executed: file_path=\'/tmp/system_design.json\' block_content=\'001|{"Implementation approach":"We will implement the 2048 game using plain JavaScript and HTML, ensuring no frameworks are used. The game logic will handle tile movements, merging, and game state updates. The UI will be simple and clean, with a responsive design to fit different screen sizes. We will use CSS for styling and ensure the game is playable with keyboard arrow keys. The game will display the current score, have a restart button, and show a game over message when no more moves are possible.","File list":["index.html","style.css","script.js"],"Data structures and interfaces":"\\\\nclassDiagram\\\\n    class Game {\\\\n        -grid: int[][]\\\\n        -score: int\\\\n        +init(): void\\\\n        +move(direction: str): void\\\\n        +merge(direction: str): void\\\\n        +isGameOver(): bool\\\\n        +restart(): void\\\\n    }\\\\n    class UI {\\\\n        -game: Game\\\\n        +init(): void\\\\n        +update(): void\\\\n        +showGameOver(): void\\\\n        +bindEvents(): void\\\\n    }\\\\n    Game --> UI\\\\n","Program call flow":"\\\\nsequenceDiagram\\\\n    participant U as UI\\\\n    participant G as Game\\\\n    U->>G: init()\\\\n    G-->>U: return\\\\n    U->>U: bindEvents()\\\\n    U->>G: move(direction)\\\\n    G->>G: merge(direction)\\\\n    G->>U: update()\\\\n    U->>U: update()\\\\n    U->>G: isGameOver()\\\\n    G-->>U: return bool\\\\n    alt Game Over\\\\n        U->>U: showGameOver()\\\\n    end\\\\n    U->>G: restart()\\\\n    G-->>U: return\\\\n","Anything UNCLEAR":"Clarify if there are any specific design preferences or additional features required beyond the basic 2048 game functionality."}\\n\'\n\nCommand Editor.read executed: file_path=\'/tmp/project_schedule.json\' block_content=\'001|{"Required packages":["No third-party dependencies required"],"Required Other language third-party packages":["No third-party dependencies required"],"Logic Analysis":[["script.js","Contains Game and UI classes, and their methods: init, move, merge, isGameOver, restart, update, showGameOver, bindEvents"],["index.html","Contains the HTML structure for the game UI"],["style.css","Contains the CSS styles for the game UI"]],"Task list":["index.html","style.css","script.js"],"Full API spec":"","Shared Knowledge":"The `script.js` file will contain the core game logic and UI handling. The `index.html` file will provide the structure for the game, and `style.css` will handle the styling.","Anything UNCLEAR":"Clarify if there are any specific design preferences or additional features required beyond the basic 2048 game functionality."}\\n\'',
-    },
-    {
-        "role": "user",
-        "content": "\n# Current Plan\n{'goal': \"Please write a 1048 game using JavaScript and HTML code without using any frameworks, user can play with keyboard. Refer to the system design located at '/tmp/system_design.json' and the project schedule at '/tmp/project_schedule.json' for detailed information.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nBased on the context, write a plan or modify an existing plan to achieve the goal. A plan consists of one to 3 tasks.\nIf plan is created, you should track the progress and update the plan accordingly, such as Plan.finish_current_task, Plan.append_task, Plan.reset_task, Plan.replace_task, etc.\nWhen presented a current task, tackle the task using the available commands.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to new user requirement.\nNote:\n1. If you keeping encountering errors, unexpected situation, or you are not sure of proceeding, use RoleZero.ask_human to ask for help.\n2. Carefully review your progress at the current task, if your actions so far has not fulfilled the task instruction, you should continue with current task. Otherwise, finish current task.\n3. Each time you finish a task, use RoleZero.reply_to_human to report your progress.\n4. Each time you write a code in your response, write with the Editor directly without preparing a repetitive code block beforehand.\n5. Take on ONE task and write ONE code file in each response. DON'T attempt all tasks in one response.\n6. When not specified, you should write files in a folder named \"src\". If you know the project path, then write in a \"src\" folder under the project path.\n7. When provided system design or project schedule, you MUST read them first before making a plan, then adhere to them in your implementation, especially in the programming language, package, or framework. You MUST implement all code files prescribed in the system design or project schedule. You can create a plan first with each task corresponding to implementing one code file.\n8. Write at most one file per task, do your best to implement THE ONLY ONE FILE. CAREFULLY CHECK THAT YOU DONT MISS ANY NECESSARY CLASS/FUNCTION IN THIS FILE.\n9. COMPLETE CODE: Your code will be part of the entire project, so please implement complete, reliable, reusable code snippets.\n10. When provided system design, YOU MUST FOLLOW \"Data structures and interfaces\". DONT CHANGE ANY DESIGN. Do not use public member functions that do not exist in your design.\n11. Write out EVERY CODE DETAIL, DON'T LEAVE TODO.\n12. To modify code in a file, read the entire file, make changes, and update the file with the complete code, ensuring that no line numbers are included in the final write.\n13. When a system design or project schedule is provided, at the end of the plan, add a CodeRview Task for each file; for example, if there are three files, add three CodeRview Tasks. For each CodeRview Task, just call ReviewAndRewriteCode.run.\n",
-    },
-]
-REQ1 = json.dumps(OBJ1)
-REQ2 = json.dumps(OBJ2)
-
-RESP1 = """
-Explanation: To begin the task of creating a 2048 game using JavaScript and HTML, I need to first read the system design and project schedule documents to understand the specific requirements and architecture outlined for the game development.
-
-```json
-[
-    {
-        "command_name": "Editor.read",
-        "args": {
-            "path": "/tmp/system_design.json"
-        }
-    },
-    {
-        "command_name": "Editor.read",
-        "args": {
-            "path": "/tmp/project_schedule.json"
-        }
-    }
-]
-```
-"""
-RESP2 = """
-Explanation: Based on the system design and project schedule, I will create a plan to implement the 2048 game. The plan will consist of tasks to create the necessary files (`index.html`, `style.css`, and `script.js`) and review each file to ensure they meet the design specifications.
-
-Here's the plan:
-
-1. **Task 1**: Create `index.html` - This file will contain the HTML structure necessary for the game's UI.
-2. **Task 2**: Create `style.css` - This file will define the CSS styles to make the game visually appealing and responsive.
-3. **Task 3**: Create `script.js` - This file will contain the JavaScript code for the game logic and UI interactions.
-4. **Code Review Tasks**: Review each file to ensure they meet the project requirements and adhere to the system design.
-
-Let's start by appending the first task to the plan.
-
-```json
-[
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "1",
-            "dependent_task_ids": [],
-            "instruction": "Create the index.html file with the basic HTML structure for the 2048 game.",
-            "assignee": "Alex"
-        }
-    },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "2",
-            "dependent_task_ids": ["1"],
-            "instruction": "Create the style.css file with the necessary CSS to style the 2048 game.",
-            "assignee": "Alex"
-        }
-    },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "3",
-            "dependent_task_ids": ["1", "2"],
-            "instruction": "Create the script.js file containing the JavaScript logic for the 2048 game.",
-            "assignee": "Alex"
-        }
-    },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "4",
-            "dependent_task_ids": ["1"],
-            "instruction": "Use ReviewAndRewriteCode to review the code in index.html to ensure it meets the design specifications.",
-            "assignee": "Alex"
-        }
-    },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "5",
-            "dependent_task_ids": ["2"],
-            "instruction": "Use ReviewAndRewriteCode to review the code in style.css to ensure it meets the design specifications.",
-            "assignee": "Alex"
-        }
-    },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "6",
-            "dependent_task_ids": ["3"],
-            "instruction": "Use ReviewAndRewriteCode to review the code in script.js to ensure it meets the design specifications.",
-            "assignee": "Alex"
-        }
-    }
-]
-```
-"""
-
-ENGINEER_EXPERIENCES = ((REQ1, RESP1), (REQ2, RESP2))
diff --git a/examples/exp_pool/init_exp_pool/main.py b/examples/exp_pool/init_exp_pool/main.py
deleted file mode 100644
index 849834513..000000000
--- a/examples/exp_pool/init_exp_pool/main.py
+++ /dev/null
@@ -1,55 +0,0 @@
-"""Init experience pool.
-
-Put some useful experiences into the experience pool.
-"""
-
-import asyncio
-
-from examples.exp_pool.init_exp_pool.engineer_exps import ENGINEER_EXPERIENCES
-from examples.exp_pool.init_exp_pool.team_leader_exps import TEAM_LEADER_EXPERIENCES
-from metagpt.exp_pool import exp_manager
-from metagpt.exp_pool.schema import EntryType, Experience, Metric, Score
-from metagpt.logs import logger
-
-
-async def add_exp(req: str, resp: str, tag: str, metric: Metric = None):
-    exp = Experience(
-        req=req,
-        resp=resp,
-        entry_type=EntryType.MANUAL,
-        tag=tag,
-        metric=metric or Metric(score=Score(val=10, reason="Manual")),
-    )
-
-    exp_manager.config.exp_pool.enable_write = True
-    exp_manager.create_exp(exp)
-    logger.info(f"New experience created for the request `{req[:10]}`.")
-
-
-async def add_teamleader_exps():
-    tag = "TeamLeader.llm_cached_aask"
-
-    for req, resp in TEAM_LEADER_EXPERIENCES:
-        await add_exp(req=req, resp=resp, tag=tag)
-
-
-async def add_engineer_exps():
-    tag = "Engineer2.llm_cached_aask"
-
-    for req, resp in ENGINEER_EXPERIENCES:
-        await add_exp(req=req, resp=resp, tag=tag)
-
-
-def query_exps_count():
-    count = exp_manager.get_exps_count()
-    logger.info(f"Experiences Count: {count}")
-
-
-async def main():
-    await add_teamleader_exps()
-    await add_engineer_exps()
-    query_exps_count()
-
-
-if __name__ == "__main__":
-    asyncio.run(main())
diff --git a/examples/exp_pool/init_exp_pool/team_leader_exps.py b/examples/exp_pool/init_exp_pool/team_leader_exps.py
deleted file mode 100644
index 347faac45..000000000
--- a/examples/exp_pool/init_exp_pool/team_leader_exps.py
+++ /dev/null
@@ -1,181 +0,0 @@
-import json
-
-OBJ1 = [
-    {
-        "role": "user",
-        "content": "\n# Current Plan\n{'goal': \"from  to {'<all>'}: Write a 1024 game using JavaScript and HTML code without using any frameworks, user can play with keyboard.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n",
-    }
-]
-OBJ2 = [
-    {
-        "role": "user",
-        "content": "\n# Current Plan\n{'goal': \"from  to {'<all>'}: Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n",
-    }
-]
-OBJ4 = [
-    {
-        "role": "user",
-        "content": "\n# Current Plan\n{'goal': \"from  to {'<all>'}: how does the project go?\", 'tasks': []}\n\n# Current Task\n\n\n# Instruction\nYou are a team leader, and you are responsible for drafting tasks and routing tasks to your team members.\nYour team member:\nTim: Team Leader, \nAlice: Product Manager, efficiently create a successful product that meets market demands and user expectations\nBob: Architect, design a concise, usable, complete software system\nEve: Project Manager, break down tasks according to PRD/technical design, generate a task list, and analyze task dependencies to start with the prerequisite modules\nAlex: Engineer, Take on game, app, and web development\nDavid: DataAnalyst, Take on any data-related tasks, such as data analysis, machine learning, deep learning, web browsing, web scraping, web searching, web deployment, terminal operation, git and github operation, etc.\n\nYou should NOT assign consecutive tasks to the same team member, instead, assign an aggregated task (or the complete requirement) and let the team member to decompose it.\nWhen creating a new plan involving multiple members, create all tasks at once.\nIf plan is created, you should track the progress based on team member feedback message, and update plan accordingly, such as Plan.finish_current_task, Plan.reset_task, Plan.replace_task, etc.\nYou should use TeamLeader.publish_team_message to team members, asking them to start their task. DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.\nPay close attention to new user message, review the conversation history, use RoleZero.reply_to_human to respond to the user directly, DON'T ask your team members.\n\nNote:\n1. If the requirement is a pure DATA-RELATED requirement, such as bug fixes, issue reporting, environment setup, terminal operations, pip install, web browsing, web scraping, web searching, web imitation, data science, data analysis, machine learning, deep learning, text-to-image etc. DON'T decompose it, assign a single task with the original user requirement as instruction directly to Data Analyst.\n2. If the requirement is developing a software, game, app, or website, excluding the above data-related tasks, you should decompose the requirement into multiple tasks and assign them to different team members based on their expertise, usually the sequence of Product Manager -> Architect -> Project Manager -> Engineer -> (optional: QaEngine if present) -> (optional: DataAnalyst if user requests deployment), each assigned ONE task. When publishing message to Product Manager, you should directly copy the full original user requirement.\n3. If the requirement contains both DATA-RELATED part mentioned in 1 and software development part mentioned in 2, you should decompose the software development part and assign them to different team members based on their expertise, and assign the DATA-RELATED part to Data Analyst David directly.\n4. If the requirement is a common-sense, logical, or math problem, you should respond directly without assigning any task to team members.\n5. If you think the requirement is not clear or ambiguous, you should ask the user for clarification immediately. Assign tasks only after all info is clear.\n6. It is helpful for Engineer to have both the system design and the project schedule for writing the code, so include paths of both files (if available) and remind Engineer to definitely read them when publishing message to Engineer.\n7. If the requirement is writing a TRD and software framework, you should assign it to Architect. When publishing message to Architect, you should directly copy the full original user requirement.\n",
-    }
-]
-
-REQ1 = json.dumps(OBJ1)
-REQ2 = json.dumps(OBJ2)
-REQ3 = """Conversation History:
-[
-    ...,
-    {'role': 'assistant', 'content': 'from Alice(Product Manager) to {'Bob'}: {'docs': {'20240424153821.json': {'root_path': 'docs/prd', 'filename': '20240424153821.json', 'content': '{"Language":"en_us","Programming Language":"Python","Original Requirements":"create a cli snake game","Project Name":"snake_game","Product Goals":["Develop an intuitive and addictive snake game",...], ...}}}}},
-]
-"""
-REQ4 = json.dumps(OBJ4)
-
-RESP1 = """
-Explanation: The requirement is about software development. Assign each tasks to a different team member based on their expertise. When publishing message to Product Manager, we copy original user requirement directly to ensure no information loss.
-```json
-[
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "1",
-            "dependent_task_ids": [],
-            "instruction": "Create a product requirement document (PRD) outlining the features, user interface, and user experience of the CLI python snake game.",
-            "assignee": "Alice"
-        }
-    },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "2",
-            "dependent_task_ids": ["1"],
-            "instruction": "Design the software architecture for the CLI snake game, including the choice of programming language, libraries, and data flow.",
-            "assignee": "Bob"
-        }
-    },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "3",
-            "dependent_task_ids": ["2"],
-            "instruction": "Break down the architecture into manageable tasks, identify task dependencies, and prepare a detailed task list for implementation.",
-            "assignee": "Eve"
-        }
-    },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "4",
-            "dependent_task_ids": ["3"],
-            "instruction": "Implement the core game logic for the CLI snake game, including snake movement, food generation, and score tracking.",
-            "assignee": "Alex"
-        }
-    },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "5",
-            "dependent_task_ids": ["4"],
-            "instruction": "Write comprehensive tests for the game logic and user interface to ensure functionality and reliability.",
-            "assignee": "Edward"
-        }
-    },
-    {
-        "command_name": "TeamLeader.publish_message",
-        "args": {
-            "content": "Create a cli snake game using Python",
-            "send_to": "Alice"
-        }
-    },
-    {
-        "command_name": "RoleZero.reply_to_human",
-        "args": {
-            "content": "I have assigned the tasks to the team members. Alice will create the PRD, Bob will design the software architecture, Eve will break down the architecture into tasks, Alex will implement the core game logic, and Edward will write comprehensive tests. The team will work on the project accordingly"
-        }
-    },
-    {
-        "command_name": "end"
-    }
-]
-```
-"""
-RESP2 = """
-Explanation: DON'T decompose requirement if it is a DATA-RELATED task, assign a single task directly to Data Analyst David. He will manage the decomposition and implementation.
-```json
-[
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "1",
-            "dependent_task_ids": [],
-            "instruction": "Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.",
-            "assignee": "David"
-        }
-    },
-    {
-        "command_name": "TeamLeader.publish_message",
-        "args": {
-            "content": "Run data analysis on sklearn Wine recognition dataset, include a plot, and train a model to predict wine class (20% as validation), and show validation accuracy.",
-            "send_to": "David"
-        }
-    },
-    {
-        "command_name": "RoleZero.reply_to_human",
-        "args": {
-            "content": "I have assigned the task to David. He will break down the task further by himself and starts solving it."
-        }
-    },
-    {
-        "command_name": "end"
-    }
-]
-```
-"""
-RESP3 = """
-Explanation: You received a message from Alice, the Product Manager, that she has completed the PRD, use Plan.finish_current_task to mark her task as finished and moves the plan to the next task. Based on plan status, next task is for Bob (Architect), publish a message asking him to start. The message content should contain important path info.
-```json
-[
-    {
-        "command_name": "Plan.finish_current_task",
-        "args": {}
-    },
-    {
-        "command_name": "TeamLeader.publish_message",
-        "args": {
-            "content": "Please design the software architecture for the snake game based on the PRD created by Alice. The PRD is at 'docs/prd/20240424153821.json'. Include the choice of programming language, libraries, and data flow, etc.",
-            "send_to": "Bob"
-        }
-    },
-    {
-        "command_name": "RoleZero.reply_to_human",
-        "args": {
-            "content": "Alice has completed the PRD. I have marked her task as finished and sent the PRD to Bob. Bob will work on the software architecture."
-        }
-    },
-    {
-        "command_name": "end"
-    }
-]
-```
-"""
-RESP4 = """
-Explanation: The user is asking for a general update on the project status. Give a straight answer about the current task the team is working on and provide a summary of the completed tasks.
-```json
-[
-    {
-        "command_name": "RoleZero.reply_to_human",
-        "args": {
-            "content": "The team is currently working on ... We have completed ..."
-        }
-    },
-    {
-        "command_name": "end"
-    }
-]
-```
-"""
-
-TEAM_LEADER_EXPERIENCES = (
-    (REQ1, RESP1),
-    (REQ2, RESP2),
-    (REQ3, RESP3),
-    (REQ4, RESP4),
-)
diff --git a/examples/exp_pool/manager.py b/examples/exp_pool/manager.py
index 3216e78b8..ae998214a 100644
--- a/examples/exp_pool/manager.py
+++ b/examples/exp_pool/manager.py
@@ -1,4 +1,8 @@
-"""Simple example of experience pool."""
+"""
+Demonstrate the creation and querying of experiences.
+
+This script creates a new experience, logs its creation, and then queries for experiences matching the same request.
+"""
 
 import asyncio
 
@@ -8,12 +12,16 @@ from metagpt.logs import logger
 
 
 async def main():
-    req = "Simple task."
-    exp = Experience(req=req, resp="echo", entry_type=EntryType.MANUAL)
+    # Define the simple request and response
+    req = "Simple req"
+    resp = "Simple resp"
 
+    # Add the new experience
+    exp = Experience(req=req, resp=resp, entry_type=EntryType.MANUAL)
     exp_manager.create_exp(exp)
     logger.info(f"New experience created for the request `{req}`.")
 
+    # Query for experiences matching the request
     exps = await exp_manager.query_exps(req)
     logger.info(f"Got experiences: {exps}")
 
diff --git a/examples/exp_pool/scorer.py b/examples/exp_pool/scorer.py
index c412feaf3..aafcee63f 100644
--- a/examples/exp_pool/scorer.py
+++ b/examples/exp_pool/scorer.py
@@ -2,13 +2,16 @@ import asyncio
 
 from metagpt.exp_pool.scorers import SimpleScorer
 
+# Request to implement quicksort in Python
 REQ = "Write a program to implement quicksort in python."
 
+# First response: Quicksort implementation without base case
 RESP1 = """
 def quicksort(arr):
     return quicksort([x for x in arr[1:] if x <= arr[0]]) + [arr[0]] + quicksort([x for x in arr[1:] if x > arr[0]])
 """
 
+# Second response: Quicksort implementation with base case
 RESP2 = """
 def quicksort(arr):
     if len(arr) <= 1:
@@ -18,6 +21,15 @@ def quicksort(arr):
 
 
 async def simple():
+    """Evaluates two quicksort implementations using SimpleScorer.
+
+    Example:
+        {
+            "val": 3,
+            "reason": "The response attempts to implement quicksort but contains a critical flaw: it lacks a base case to terminate the recursion, which will lead to a maximum recursion depth exceeded error for non-empty lists. Additionally, the function does not handle empty lists properly. A correct implementation should include a base case to handle lists of length 0 or 1."
+        }
+    """
+
     scorer = SimpleScorer()
 
     await scorer.evaluate(req=REQ, resp=RESP1)
diff --git a/tests/metagpt/exp_pool/test_manager.py b/tests/metagpt/exp_pool/test_manager.py
index 2f712ad44..3fb2aaa6d 100644
--- a/tests/metagpt/exp_pool/test_manager.py
+++ b/tests/metagpt/exp_pool/test_manager.py
@@ -3,9 +3,8 @@ import pytest
 from metagpt.config2 import Config
 from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 from metagpt.configs.llm_config import LLMConfig
-from metagpt.exp_pool.manager import ExperienceManager
-from metagpt.exp_pool.schema import Experience
-from metagpt.rag.engines import SimpleEngine
+from metagpt.exp_pool.manager import Experience, ExperienceManager
+from metagpt.exp_pool.schema import QueryType
 
 
 class TestExperienceManager:
@@ -15,50 +14,65 @@ class TestExperienceManager:
 
     @pytest.fixture
     def mock_storage(self, mocker):
-        engine = mocker.MagicMock(spec=SimpleEngine)
+        engine = mocker.MagicMock()
         engine.add_objs = mocker.MagicMock()
         engine.aretrieve = mocker.AsyncMock(return_value=[])
         engine._retriever = mocker.MagicMock()
         engine._retriever._vector_store = mocker.MagicMock()
-        engine._retriever._vector_store._get = mocker.MagicMock(return_value=mocker.MagicMock(ids=[]))
+        engine._retriever._vector_store._collection = mocker.MagicMock()
+        engine._retriever._vector_store._collection.count = mocker.MagicMock(return_value=10)
         return engine
 
     @pytest.fixture
-    def mock_experience_manager(self, mock_config, mock_storage):
-        return ExperienceManager(config=mock_config, storage=mock_storage)
+    def exp_manager(self, mock_config, mock_storage):
+        manager = ExperienceManager(config=mock_config)
+        manager._storage = mock_storage
+        return manager
 
-    @pytest.fixture
-    def mock_experience(self):
-        return Experience(req="req", resp="resp")
-
-    def test_initialize_storage(self, mock_experience_manager, mock_storage):
-        assert mock_experience_manager.storage is mock_storage
-
-    def test_create_exp(self, mock_experience_manager, mock_experience):
-        mock_experience_manager.create_exp(mock_experience)
-        mock_experience_manager.storage.add_objs.assert_called_with([mock_experience])
-
-    def test_create_exp_write_disabled(self, mock_experience_manager, mock_experience, mock_config):
-        mock_config.exp_pool.enable_write = False
-        mock_experience_manager.create_exp(mock_experience)
-        mock_experience_manager.storage.add_objs.assert_not_called()
+    def test_vector_store_property(self, exp_manager):
+        assert exp_manager.vector_store == exp_manager.storage._retriever._vector_store
 
     @pytest.mark.asyncio
-    async def test_query_exps(self, mock_experience_manager, mocker):
-        req = "req"
-        resp = "resp"
-        tag = "test"
-        experiences = [Experience(req=req, resp=resp, tag="test"), Experience(req=req, resp=resp, tag="other")]
-        mock_experience_manager.storage.aretrieve.return_value = [
-            mocker.MagicMock(metadata={"obj": exp}) for exp in experiences
-        ]
+    async def test_query_exps_with_exact_match(self, exp_manager, mocker):
+        req = "exact query"
+        exp1 = Experience(req=req, resp="response1")
+        exp2 = Experience(req="different query", resp="response2")
 
-        result = await mock_experience_manager.query_exps(req, tag)
+        mock_node1 = mocker.MagicMock(metadata={"obj": exp1})
+        mock_node2 = mocker.MagicMock(metadata={"obj": exp2})
+
+        exp_manager.storage.aretrieve.return_value = [mock_node1, mock_node2]
+
+        result = await exp_manager.query_exps(req, query_type=QueryType.EXACT)
         assert len(result) == 1
-        assert result[0].tag == "test"
+        assert result[0].req == req
 
     @pytest.mark.asyncio
-    async def test_query_exps_no_read_permission(self, mock_experience_manager, mock_config):
+    async def test_query_exps_with_tag_filter(self, exp_manager, mocker):
+        tag = "test_tag"
+        exp1 = Experience(req="query1", resp="response1", tag=tag)
+        exp2 = Experience(req="query2", resp="response2", tag="other_tag")
+
+        mock_node1 = mocker.MagicMock(metadata={"obj": exp1})
+        mock_node2 = mocker.MagicMock(metadata={"obj": exp2})
+
+        exp_manager.storage.aretrieve.return_value = [mock_node1, mock_node2]
+
+        result = await exp_manager.query_exps("query", tag=tag)
+        assert len(result) == 1
+        assert result[0].tag == tag
+
+    def test_get_exps_count(self, exp_manager):
+        assert exp_manager.get_exps_count() == 10
+
+    def test_create_exp_write_disabled(self, exp_manager, mock_config):
+        mock_config.exp_pool.enable_write = False
+        exp = Experience(req="test", resp="response")
+        exp_manager.create_exp(exp)
+        exp_manager.storage.add_objs.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_query_exps_read_disabled(self, exp_manager, mock_config):
         mock_config.exp_pool.enable_read = False
-        result = await mock_experience_manager.query_exps("query")
+        result = await exp_manager.query_exps("query")
         assert result == []
diff --git a/tests/metagpt/exp_pool/test_serializers/test_role_zero.py b/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
index d4525d535..a168650fc 100644
--- a/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
+++ b/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
@@ -22,7 +22,7 @@ class TestRoleZeroSerializer:
         return [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]
 
     def test_serialize_req_empty_input(self, serializer: RoleZeroSerializer):
-        assert serializer.serialize_req([]) == ""
+        assert serializer.serialize_req(req=[]) == ""
 
     def test_serialize_req_with_content(self, serializer: RoleZeroSerializer, last_item: dict):
         req = [
@@ -33,7 +33,7 @@ class TestRoleZeroSerializer:
         expected_output = json.dumps(
             [{"role": "user", "content": "Command Editor.read executed: file_path=test.py"}, last_item]
         )
-        assert serializer.serialize_req(req) == expected_output
+        assert serializer.serialize_req(req=req) == expected_output
 
     def test_filter_req(self, serializer: RoleZeroSerializer):
         req = [
diff --git a/tests/metagpt/exp_pool/test_serializers/test_simple.py b/tests/metagpt/exp_pool/test_serializers/test_simple.py
index 05ef1ca11..2a6bf96e3 100644
--- a/tests/metagpt/exp_pool/test_serializers/test_simple.py
+++ b/tests/metagpt/exp_pool/test_serializers/test_simple.py
@@ -8,28 +8,28 @@ class TestSimpleSerializer:
     def serializer(self):
         return SimpleSerializer()
 
-    def test_serialize_req(self, serializer):
+    def test_serialize_req(self, serializer: SimpleSerializer):
         # Test with different types of input
-        assert serializer.serialize_req(123) == "123"
-        assert serializer.serialize_req("test") == "test"
-        assert serializer.serialize_req([1, 2, 3]) == "[1, 2, 3]"
-        assert serializer.serialize_req({"a": 1}) == "{'a': 1}"
+        assert serializer.serialize_req(req=123) == "123"
+        assert serializer.serialize_req(req="test") == "test"
+        assert serializer.serialize_req(req=[1, 2, 3]) == "[1, 2, 3]"
+        assert serializer.serialize_req(req={"a": 1}) == "{'a': 1}"
 
-    def test_serialize_resp(self, serializer):
+    def test_serialize_resp(self, serializer: SimpleSerializer):
         # Test with different types of input
         assert serializer.serialize_resp(456) == "456"
         assert serializer.serialize_resp("response") == "response"
         assert serializer.serialize_resp([4, 5, 6]) == "[4, 5, 6]"
         assert serializer.serialize_resp({"b": 2}) == "{'b': 2}"
 
-    def test_deserialize_resp(self, serializer):
+    def test_deserialize_resp(self, serializer: SimpleSerializer):
         # Test with different types of input
         assert serializer.deserialize_resp("789") == "789"
         assert serializer.deserialize_resp("test_response") == "test_response"
         assert serializer.deserialize_resp("[7, 8, 9]") == "[7, 8, 9]"
         assert serializer.deserialize_resp("{'c': 3}") == "{'c': 3}"
 
-    def test_roundtrip(self, serializer):
+    def test_roundtrip(self, serializer: SimpleSerializer):
         # Test serialization and deserialization roundtrip
         original = "test_roundtrip"
         serialized = serializer.serialize_resp(original)
@@ -37,8 +37,8 @@ class TestSimpleSerializer:
         assert deserialized == original
 
     @pytest.mark.parametrize("input_value", [123, "test", [1, 2, 3], {"a": 1}, None])
-    def test_serialize_req_types(self, serializer, input_value):
+    def test_serialize_req_types(self, serializer: SimpleSerializer, input_value):
         # Test serialize_req with various input types
-        result = serializer.serialize_req(input_value)
+        result = serializer.serialize_req(req=input_value)
         assert isinstance(result, str)
         assert result == str(input_value)

From 4e1955eba8199cdcdee890f26b24d7b8bfcb427b Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 16 Jul 2024 20:52:44 +0800
Subject: [PATCH 49/51] update comment

---
 metagpt/exp_pool/context_builders/role_zero.py | 10 +++++-----
 metagpt/prompts/di/role_zero.py                |  6 +++---
 2 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/metagpt/exp_pool/context_builders/role_zero.py b/metagpt/exp_pool/context_builders/role_zero.py
index aa5524ab4..924bd56aa 100644
--- a/metagpt/exp_pool/context_builders/role_zero.py
+++ b/metagpt/exp_pool/context_builders/role_zero.py
@@ -12,12 +12,12 @@ class RoleZeroContextBuilder(BaseContextBuilder):
         """Builds the role zero context string.
 
         Note:
-            1. The expected format for `req`, e.g., [{...}, {"role": "user", "content": "context"}, {"role": "user", "content": "context exp part"}].
-            2. Returns the original `req` if it is empty, incorrectly formatted or there are no experiences.
+            1. The expected format for `req`, e.g., [{...}, {"role": "user", "content": "context"}].
+            2. Returns the original `req` if it is empty.
             3. Creates a copy of req and replaces the example content in the copied req with actual experiences.
         """
 
-        if not req or len(req) < 2:
+        if not req:
             return req
 
         exps = self.format_exps()
@@ -26,12 +26,12 @@ class RoleZeroContextBuilder(BaseContextBuilder):
 
         req_copy = copy.deepcopy(req)
 
-        req_copy[-2]["content"] = self.replace_example_content(req_copy[-2].get("content", ""), exps)
+        req_copy[-1]["content"] = self.replace_example_content(req_copy[-1].get("content", ""), exps)
 
         return req_copy
 
     def replace_example_content(self, text: str, new_example_content: str) -> str:
-        return self.replace_content_between_markers(text, "# Example", "# Available Commands", new_example_content)
+        return self.replace_content_between_markers(text, "# Example", "# Instruction", new_example_content)
 
     @staticmethod
     def replace_content_between_markers(text: str, start_marker: str, end_marker: str, new_content: str) -> str:
diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index 41b9e023e..b2e931e23 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -18,9 +18,6 @@ class Task(BaseModel):
     task_type: str = ""
     assignee: str = ""
 
-# Example
-{example}
-
 # Available Commands
 {available_commands}
 Special Command: Use {{"command_name": "end"}} to do nothing or indicate completion of all requirements and the end of actions.
@@ -34,6 +31,9 @@ Special Command: Use {{"command_name": "end"}} to do nothing or indicate complet
 # Current Task
 {current_task}
 
+# Example
+{example}
+
 # Instruction
 {instruction}
 

From d612d826d52f90eda32b745bb1cc2ebc4242f2d8 Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 16 Jul 2024 21:13:16 +0800
Subject: [PATCH 50/51] update comment

---
 metagpt/prompts/di/role_zero.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index b2e931e23..890c1d562 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -17,14 +17,14 @@ class Task(BaseModel):
     instruction: str = ""
     task_type: str = ""
     assignee: str = ""
+    
+# Available Task Types
+{task_type_desc}
 
 # Available Commands
 {available_commands}
 Special Command: Use {{"command_name": "end"}} to do nothing or indicate completion of all requirements and the end of actions.
 
-# Available Task Types
-{task_type_desc}
-
 # Current Plan
 {plan_status}
 

From e70a08045460cedab612849c2a5b6ff9b92269eb Mon Sep 17 00:00:00 2001
From: seehi <6580@pm.me>
Date: Tue, 16 Jul 2024 21:20:01 +0800
Subject: [PATCH 51/51] update exp_pool tests

---
 .../test_rolezero_context_builder.py           | 18 +++++++-----------
 .../test_serializers/test_role_zero.py         |  4 +---
 2 files changed, 8 insertions(+), 14 deletions(-)

diff --git a/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py b/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
index a95566ed1..b7182602d 100644
--- a/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
+++ b/tests/metagpt/exp_pool/test_context_builders/test_rolezero_context_builder.py
@@ -25,31 +25,27 @@ class TestRoleZeroContextBuilder:
     async def test_build_with_experiences(self, context_builder, mocker):
         mocker.patch.object(BaseContextBuilder, "format_exps", return_value="Formatted experiences")
         mocker.patch.object(RoleZeroContextBuilder, "replace_example_content", return_value="Updated content")
-        req = [{"content": "Original content 1"}, {"content": "Original content exp part"}]
+        req = [{"content": "Original content 1"}]
         result = await context_builder.build(req=req)
-        assert result == [{"content": "Updated content"}, {"content": "Original content exp part"}]
+        assert result == [{"content": "Updated content"}]
 
     def test_replace_example_content(self, context_builder, mocker):
         mocker.patch.object(RoleZeroContextBuilder, "replace_content_between_markers", return_value="Replaced content")
         result = context_builder.replace_example_content("Original text", "New example content")
         assert result == "Replaced content"
         context_builder.replace_content_between_markers.assert_called_once_with(
-            "Original text", "# Example", "# Available Commands", "New example content"
+            "Original text", "# Example", "# Instruction", "New example content"
         )
 
     def test_replace_content_between_markers(self):
-        text = "Start\n# Example\nOld content\n# Available Commands\nEnd"
+        text = "Start\n# Example\nOld content\n# Instruction\nEnd"
         new_content = "New content"
-        result = RoleZeroContextBuilder.replace_content_between_markers(
-            text, "# Example", "# Available Commands", new_content
-        )
-        expected = "Start\n# Example\nNew content\n\n# Available Commands\nEnd"
+        result = RoleZeroContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
+        expected = "Start\n# Example\nNew content\n\n# Instruction\nEnd"
         assert result == expected
 
     def test_replace_content_between_markers_no_match(self):
         text = "Start\nNo markers\nEnd"
         new_content = "New content"
-        result = RoleZeroContextBuilder.replace_content_between_markers(
-            text, "# Example", "# Available Commands", new_content
-        )
+        result = RoleZeroContextBuilder.replace_content_between_markers(text, "# Example", "# Instruction", new_content)
         assert result == text
diff --git a/tests/metagpt/exp_pool/test_serializers/test_role_zero.py b/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
index a168650fc..964443f29 100644
--- a/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
+++ b/tests/metagpt/exp_pool/test_serializers/test_role_zero.py
@@ -30,9 +30,7 @@ class TestRoleZeroSerializer:
             {"role": "assistant", "content": "Some other content"},
             last_item,
         ]
-        expected_output = json.dumps(
-            [{"role": "user", "content": "Command Editor.read executed: file_path=test.py"}, last_item]
-        )
+        expected_output = json.dumps([{"role": "user", "content": "Command Editor.read executed: file_path=test.py"}])
         assert serializer.serialize_req(req=req) == expected_output
 
     def test_filter_req(self, serializer: RoleZeroSerializer):