Merge pull request #1195 from luxiangtaoya/code_interpreter

add experience

metagpt/actions/di/use_experience.py

@@ -0,0 +1,225 @@
+import json
+
+import chromadb
+from pydantic import BaseModel
+
+from metagpt.actions import Action
+from metagpt.const import SERDESER_PATH
+from metagpt.logs import logger
+from metagpt.prompts.di.get_task_summary import TASK_CODE_DESCRIPTION_PROMPT
+from metagpt.rag.engines import SimpleEngine
+from metagpt.rag.schema import ChromaRetrieverConfig
+from metagpt.schema import Task
+from metagpt.strategy.planner import Planner
+
+
+class Trajectory(BaseModel):
+    user_requirement: str = ""
+    task_map: dict[str, Task] = {}
+    task: Task = None
+    is_used: bool = False
+
+    def rag_key(self) -> str:
+        """For search"""
+        return self.task.instruction
+
+
+class Experience(BaseModel):
+    code_summary: str = ""
+    trajectory: Trajectory = None
+
+    def rag_key(self) -> str:
+        """For search"""
+        return self.code_summary
+
+
+EXPERIENCE_COLLECTION_NAME = "di_experience_0"
+TRAJECTORY_COLLECTION_NAME = "di_trajectory_0"
+PERSIST_PATH = SERDESER_PATH / "data_interpreter/chroma"
+
+
+class AddNewTrajectories(Action):
+    """Record the execution status of each task as a trajectory and store it."""
+
+    name: str = "AddNewTrajectories"
+
+    def _init_engine(self, collection_name: str):
+        """Initialize a collection for storing code experiences."""
+
+        engine = SimpleEngine.from_objs(
+            retriever_configs=[ChromaRetrieverConfig(persist_path=PERSIST_PATH, collection_name=collection_name)],
+        )
+        return engine
+
+    async def run(self, planner: Planner, trajectory_collection_name: str = TRAJECTORY_COLLECTION_NAME):
+        """Initiate a collection and add new trajectories to the collection."""
+
+        engine = self._init_engine(trajectory_collection_name)
+
+        if not planner.plan.tasks:
+            return
+
+        user_requirement = planner.plan.goal
+        task_map = planner.plan.task_map
+        trajectories = [
+            Trajectory(user_requirement=user_requirement, task_map=task_map, task=task, is_used=False)
+            for task in planner.plan.tasks
+        ]
+
+        engine.add_objs(trajectories)
+
+
+class AddNewExperiences(Action):
+    """Retrieve the trajectories from the vector database where trajectories are stored,
+    compare and summarize them to form experiences, and then store these experiences in the vector database.
+    """
+
+    name: str = "AddNewTaskExperiences"
+
+    def _init_engine(self, collection_name: str):
+        """Initialize a collection for storing code experiences."""
+
+        engine = SimpleEngine.from_objs(
+            retriever_configs=[ChromaRetrieverConfig(persist_path=PERSIST_PATH, collection_name=collection_name)],
+        )
+        return engine
+
+    async def _single_task_summary(self, trajectory_collection_name: str, experience_collection_name: str):
+        trajectory_engine = self._init_engine(collection_name=trajectory_collection_name)
+        experience_engine = self._init_engine(collection_name=experience_collection_name)
+
+        db = chromadb.PersistentClient(path=str(PERSIST_PATH))
+        collection = db.get_or_create_collection(trajectory_collection_name)
+
+        # get the ids of all trajectories where the is_used attribute is false.
+        unused_ids = [
+            id
+            for id in collection.get()["ids"]  # collection.get()["ids"] will get all the ids in the collection
+            if json.loads(collection.get([id])["metadatas"][0]["obj_json"])["is_used"]
+            == False  # Check if the is_used attribute of the trajectory corresponding to the given id is false.
+        ]
+
+        trajectory_dicts = [
+            json.loads(metadata["obj_json"]) for metadata in collection.get(unused_ids)["metadatas"]
+        ]  # get the trajectory in dictionary format
+        trajectories = []
+        experiences = []
+
+        for trajectory_dict in trajectory_dicts:
+            # set the is_used attribute of the trajectory to true and create a new trajectory (the old trajectory will be deleted below).
+            trajectory_dict["is_used"] = True
+            trajectory = Trajectory(**trajectory_dict)
+            trajectories.append(trajectory)
+
+            # summarize the trajectory using LLM and assemble it into a single experience
+            code_summary = await self.task_code_sumarization(trajectory)
+            experience = Experience(code_summary=code_summary, trajectory=trajectory)
+            experiences.append(experience)
+
+        collection.delete(unused_ids)  # delete the old trajectories
+        trajectory_engine.add_objs(trajectories)
+        experience_engine.add_objs(experiences)
+
+    async def task_code_sumarization(self, trajectory: Trajectory):
+        """use LLM to summarize the task code.
+        Args:
+            trajectory: The trajectory to be summarized.
+        Returns:
+            A summary of the trajectory's code.
+        """
+        task = trajectory.task
+        prompt = TASK_CODE_DESCRIPTION_PROMPT.format(
+            code_snippet=task.code, code_result=task.result, code_success="Success" if task.is_success else "Failure"
+        )
+        resp = await self._aask(prompt=prompt)
+        return resp
+
+    async def run(
+        self,
+        trajectory_collection_name: str = TRAJECTORY_COLLECTION_NAME,
+        experience_collection_name: str = EXPERIENCE_COLLECTION_NAME,
+        mode: str = "single_task_summary",
+    ):
+        """Initiate a collection and Add a new task experience to the collection.
+
+        Args:
+            trajectory_collection_name(str): the trajectory collection_name to be used for geting experiences.
+            experience_collection_name(str): the experience collection_name to be used for saving experiences.
+            mode(str): how to generate experiences.
+
+        """
+        if mode == "single_task_summary":
+            await self._single_task_summary(
+                trajectory_collection_name=trajectory_collection_name,
+                experience_collection_name=experience_collection_name,
+            )
+        else:
+            pass  # TODO:add other methods to generate experiences from trajectories.
+
+
+class RetrieveExperiences(Action):
+    """Retrieve the most relevant experience from the vector database based on the input task."""
+
+    name: str = "RetrieveExperiences"
+
+    def _init_engine(self, collection_name: str, top_k: int):
+        """Initialize a SimpleEngine for retrieving experiences.
+
+        Args:
+            query (str): The chromadb collectin_name.
+            top_k (int): The number of eperiences to be retrieved.
+        """
+
+        engine = SimpleEngine.from_objs(
+            retriever_configs=[
+                ChromaRetrieverConfig(
+                    persist_path=PERSIST_PATH, collection_name=collection_name, similarity_top_k=top_k
+                )
+            ],
+        )
+        return engine
+
+    async def run(
+        self, query: str, experience_collection_name: str = EXPERIENCE_COLLECTION_NAME, top_k: int = 5
+    ) -> str:
+        """Retrieve past attempted tasks
+
+        Args:
+            query (str): The task instruction to be used for retrieval.
+            experience_collection_name(str): the collextion_name for retrieving experiences.
+            top_k (int, optional): The number of experiences to be retrieved. Defaults to 5.
+
+        Returns:
+            _type_: _description_
+        """
+        engine = self._init_engine(collection_name=experience_collection_name, top_k=top_k)
+
+        if len(query) <= 2:  # not "" or not '""'
+            return ""
+
+        nodes = await engine.aretrieve(query)
+        new_experiences = []
+        for i, node in enumerate(nodes):
+            try:
+                code_summary = node.node.metadata["obj"].code_summary
+                trajectory = node.node.metadata["obj"].trajectory
+
+            except:
+                continue
+
+            # Create the experience dictionary with placeholder keys
+            experience = {
+                "Reference __i__": trajectory.task.instruction,
+                "Task code": trajectory.task.code,
+                "Code summary": code_summary,
+                "Task result": trajectory.task.result,
+                "Task outcome": "Success" if trajectory.task.is_success else "Failure",
+                "Task ownership's requirement": "This task is part of " + trajectory.user_requirement,
+            }
+
+            # Replace the placeholder in the keys
+            experience = {k.replace("__i__", str(i)): v for k, v in experience.items()}
+            new_experiences.append(experience)
+
+        logger.info("retrieval done")
+        return json.dumps(new_experiences, indent=4)

metagpt/actions/di/write_analysis_code.py

@@ -41,11 +41,13 @@ class WriteAnalysisCode(Action):
         tool_info: str = "",
         working_memory: list[Message] = None,
         use_reflection: bool = False,
+        experiences: str = "",
         **kwargs,
     ) -> str:
         structual_prompt = STRUCTUAL_PROMPT.format(
             user_requirement=user_requirement,
             plan_status=plan_status,
+            experiences=experiences,
             tool_info=tool_info,
         )
 

metagpt/prompts/di/get_task_summary.py

@@ -0,0 +1,10 @@
+TASK_CODE_DESCRIPTION_PROMPT = """
+Please explain in a paragraph what the following code snippet does. Only the function of the code snippet needs to be explained, no variable names need to be explained.
+
+Code snippet:
+{code_snippet}
+Code Execution Result:
+{code_result}
+Code Success or Failure:
+{code_success}
+"""

metagpt/prompts/di/write_analysis_code.py

@@ -7,6 +7,10 @@ STRUCTUAL_PROMPT = """
 # Plan Status
 {plan_status}
 
+# Reference experience (can be empty):
+This is some previous coding experience that is similar to the current task. You can learn from the successful code and avoid the mistakes from the failed code. If there are other codes you don't know about in the experience, please don't refer to it.
+{experiences}
+
 # Tool Info
 {tool_info}
 

metagpt/roles/di/data_interpreter.py

@@ -7,6 +7,7 @@ from pydantic import Field, model_validator
 
 from metagpt.actions.di.ask_review import ReviewConst
 from metagpt.actions.di.execute_nb_code import ExecuteNbCode
+from metagpt.actions.di.use_experience import AddNewTrajectories, RetrieveExperiences
 from metagpt.actions.di.write_analysis_code import CheckData, WriteAnalysisCode
 from metagpt.logs import logger
 from metagpt.prompts.di.write_analysis_code import DATA_INFO
@@ -38,6 +39,7 @@ class DataInterpreter(Role):
     auto_run: bool = True
     use_plan: bool = True
     use_reflection: bool = False
+    use_experience: bool = False
     execute_code: ExecuteNbCode = Field(default_factory=ExecuteNbCode, exclude=True)
     tools: list[str] = []  # Use special symbol ["<all>"] to indicate use of all registered tools
     tool_recommender: ToolRecommender = None
@@ -88,6 +90,9 @@ class DataInterpreter(Role):
     async def _plan_and_act(self) -> Message:
         try:
             rsp = await super()._plan_and_act()
+            await AddNewTrajectories().run(
+                self.planner
+            )  # extract trajectories based on the execution status of each task in the planner
             await self.execute_code.terminate()
             return rsp
         except Exception as e:
@@ -96,11 +101,13 @@ class DataInterpreter(Role):
 
     async def _act_on_task(self, current_task: Task) -> TaskResult:
         """Useful in 'plan_and_act' mode. Wrap the output in a TaskResult for review and confirmation."""
-        code, result, is_success = await self._write_and_exec_code()
+        # retrieve past tasks for this task
+        experiences = await RetrieveExperiences().run(query=current_task.instruction) if self.use_experience else ""
+        code, result, is_success = await self._write_and_exec_code(experiences=experiences)
         task_result = TaskResult(code=code, result=result, is_success=is_success)
         return task_result
 
-    async def _write_and_exec_code(self, max_retry: int = 3):
+    async def _write_and_exec_code(self, max_retry: int = 3, experiences: str = ""):
         counter = 0
         success = False
 
@@ -122,7 +129,9 @@ class DataInterpreter(Role):
 
         while not success and counter < max_retry:
             ### write code ###
-            code, cause_by = await self._write_code(counter, plan_status, tool_info)
+            code, cause_by = await self._write_code(
+                counter, plan_status, tool_info, experiences=experiences if counter == 0 else ""
+            )
 
             self.working_memory.add(Message(content=code, role="assistant", cause_by=cause_by))
 
@@ -143,12 +152,7 @@ class DataInterpreter(Role):
 
         return code, result, success
 
-    async def _write_code(
-        self,
-        counter: int,
-        plan_status: str = "",
-        tool_info: str = "",
-    ):
+    async def _write_code(self, counter: int, plan_status: str = "", tool_info: str = "", experiences: str = ""):
         todo = self.rc.todo  # todo is WriteAnalysisCode
         logger.info(f"ready to {todo.name}")
         use_reflection = counter > 0 and self.use_reflection  # only use reflection after the first trial
@@ -161,6 +165,7 @@ class DataInterpreter(Role):
             tool_info=tool_info,
             working_memory=self.working_memory.get(),
             use_reflection=use_reflection,
+            experiences=experiences,
         )
 
         return code, todo

metagpt/roles/role.py

@@ -34,7 +34,7 @@ from metagpt.context_mixin import ContextMixin
 from metagpt.logs import logger
 from metagpt.memory import Memory
 from metagpt.provider import HumanProvider
-from metagpt.schema import Message, MessageQueue, SerializationMixin
+from metagpt.schema import Message, MessageQueue, SerializationMixin, Task, TaskResult
 from metagpt.strategy.planner import Planner
 from metagpt.utils.common import any_to_name, any_to_str, role_raise_decorator
 from metagpt.utils.project_repo import ProjectRepo