diff --git a/metagpt/actions/talk_action.py b/metagpt/actions/talk_action.py index 83504b62d..81caef013 100644 --- a/metagpt/actions/talk_action.py +++ b/metagpt/actions/talk_action.py @@ -34,19 +34,20 @@ class TalkAction(Action): f"maintaining the character's persona and habits. When faced with unrelated questions, playfully " f"decline to answer without revealing your AI nature to preserve the character's image.\n\n" ) - prompt += f"Background knowledge:\n{self._knowledge}\n\n" if self._knowledge else "" + prompt += f"Knowledge:\n{self._knowledge}\n\n" if self._knowledge else "" prompt += f"{self._history_summary}\n\n" - if self._history_summary != "": - prompt += "According to the historical conversation above, " + prompt += ( + "If the information is insufficient, you can search in the historical conversation or knowledge above.\n" + ) language = CONFIG.language or DEFAULT_LANGUAGE prompt += ( - f"Answer the following questions in {language}, and the answers must follow the Markdown format.\n " + f"Answer the following questions strictly in {language}, and the answers must follow the Markdown format.\n " f"{self._talk}" ) return prompt @property - def formation_prompt(self): + def prompt_bad(self): kvs = { "{role}": CONFIG.agent_description or "", "{history}": self._history_summary or "", @@ -57,6 +58,7 @@ class TalkAction(Action): prompt = TalkAction.__FORMATION_LOOSE__ for k, v in kvs.items(): prompt = prompt.replace(k, v) + logger.info(f"PROMPT: {prompt}") return prompt async def run(self, *args, **kwargs) -> ActionOutput: @@ -71,8 +73,11 @@ class TalkAction(Action): "[HISTORY_BEGIN]" and "[HISTORY_END]" tags enclose the historical conversation; "[KNOWLEDGE_BEGIN]" and "[KNOWLEDGE_END]" tags enclose the knowledge may help for your responses; "Statement" defines the work detail you need to complete at this stage; - "[ASK_BEGIN]" and [ASK_END] tags enclose the requirements for your to respond; + "[ASK_BEGIN]" and [ASK_END] tags enclose the questions; "Constraint" defines the conditions that your responses must comply with. + "Personality" defines your language style。 + "Insight" provides a deeper understanding of the characters' inner traits. + "Initial" defines the initial setup of a character. Capacity and role: {role} Statement: Your responses should align with the role-play agreement, maintaining the @@ -80,46 +85,56 @@ Statement: Your responses should align with the role-play agreement, maintaining your AI nature to preserve the character's image. [HISTORY_BEGIN] + {history} + [HISTORY_END] [KNOWLEDGE_BEGIN] + {knowledge} + [KNOWLEDGE_END] Statement: If the information is insufficient, you can search in the historical conversation or knowledge. -Statement: Answer the following questions in {language}, and the answers must follow the Markdown format - , excluding any tag likes "[HISTORY_BEGIN]", "[HISTORY_END]", "[KNOWLEDGE_BEGIN]", "[KNOWLEDGE_END]", "[ASK_BEGIN]" - , "[ASK_END]" +Statement: Unless you are a language professional, answer the following questions strictly in {language} +, and the answers must follow the Markdown format. Strictly excluding any tag likes "[HISTORY_BEGIN]" +, "[HISTORY_END]", "[KNOWLEDGE_BEGIN]", "[KNOWLEDGE_END]" in responses. -[ASK_BEGIN] + {ask} -[ASK_END]""" +""" __FORMATION_LOOSE__ = """Formation: "Capacity and role" defines the role you are currently playing; "[HISTORY_BEGIN]" and "[HISTORY_END]" tags enclose the historical conversation; "[KNOWLEDGE_BEGIN]" and "[KNOWLEDGE_END]" tags enclose the knowledge may help for your responses; "Statement" defines the work detail you need to complete at this stage; - "[ASK_BEGIN]" and [ASK_END] tags enclose the requirements for your to respond; "Constraint" defines the conditions that your responses must comply with. + "Personality" defines your language style。 + "Insight" provides a deeper understanding of the characters' inner traits. + "Initial" defines the initial setup of a character. Capacity and role: {role} Statement: Your responses should maintaining the character's persona and habits. When faced with unrelated questions , playfully decline to answer without revealing your AI nature to preserve the character's image. [HISTORY_BEGIN] + {history} + [HISTORY_END] [KNOWLEDGE_BEGIN] + {knowledge} + [KNOWLEDGE_END] Statement: If the information is insufficient, you can search in the historical conversation or knowledge. -Statement: Answer the following questions in {language}, and the answers must follow the Markdown format - , excluding any tag likes "[HISTORY_BEGIN]", "[HISTORY_END]", "[KNOWLEDGE_BEGIN]", "[KNOWLEDGE_END]", "[ASK_BEGIN]" - , "[ASK_END]" +Statement: Unless you are a language professional, answer the following questions strictly in {language} +, and the answers must follow the Markdown format. Strictly excluding any tag likes "[HISTORY_BEGIN]" +, "[HISTORY_END]", "[KNOWLEDGE_BEGIN]", "[KNOWLEDGE_END]" in responses. + -[ASK_BEGIN] {ask} -[ASK_END]""" +""" diff --git a/metagpt/document_store/faiss_store.py b/metagpt/document_store/faiss_store.py index fbfcb3086..be4748b50 100644 --- a/metagpt/document_store/faiss_store.py +++ b/metagpt/document_store/faiss_store.py @@ -4,7 +4,6 @@ @Time : 2023/5/25 10:20 @Author : alexanderwu @File : faiss_store.py -@Modified By: mashenquan, 2023/8/20. Remove global configuration `CONFIG`, enable configuration support for business isolation. """ import pickle from pathlib import Path @@ -21,9 +20,10 @@ from metagpt.logs import logger class FaissStore(LocalStore): - def __init__(self, raw_data: Path, cache_dir=None, meta_col='source', content_col='output'): + def __init__(self, raw_data: Path, cache_dir=None, meta_col="source", content_col="output", embedding_conf=None): self.meta_col = meta_col self.content_col = content_col + self.embedding_conf = embedding_conf or {} super().__init__(raw_data, cache_dir) def _load(self) -> Optional["FaissStore"]: @@ -37,11 +37,8 @@ class FaissStore(LocalStore): store.index = index return store - def _write(self, docs, metadatas, **kwargs): - store = FAISS.from_texts(docs, - OpenAIEmbeddings(openai_api_version="2020-11-07", - openai_api_key=kwargs.get("OPENAI_API_KEY")), - metadatas=metadatas) + def _write(self, docs, metadatas): + store = FAISS.from_texts(docs, OpenAIEmbeddings(openai_api_version="2020-11-07", **self.embedding_conf), metadatas=metadatas) return store def persist(self): @@ -54,7 +51,7 @@ class FaissStore(LocalStore): pickle.dump(store, f) store.index = index - def search(self, query, expand_cols=False, sep='\n', *args, k=5, **kwargs): + def search(self, query, expand_cols=False, sep="\n", *args, k=5, **kwargs): rsp = self.store.similarity_search(query, k=k, **kwargs) logger.debug(rsp) if expand_cols: @@ -82,8 +79,8 @@ class FaissStore(LocalStore): raise NotImplementedError -if __name__ == '__main__': - faiss_store = FaissStore(DATA_PATH / 'qcs/qcs_4w.json') - logger.info(faiss_store.search('油皮洗面奶')) - faiss_store.add([f'油皮洗面奶-{i}' for i in range(3)]) - logger.info(faiss_store.search('油皮洗面奶')) +if __name__ == "__main__": + faiss_store = FaissStore(DATA_PATH / "qcs/qcs_4w.json") + logger.info(faiss_store.search("油皮洗面奶")) + faiss_store.add([f"油皮洗面奶-{i}" for i in range(3)]) + logger.info(faiss_store.search("油皮洗面奶")) diff --git a/metagpt/memory/brain_memory.py b/metagpt/memory/brain_memory.py index 586285e4f..2195da566 100644 --- a/metagpt/memory/brain_memory.py +++ b/metagpt/memory/brain_memory.py @@ -34,7 +34,7 @@ class BrainMemory(pydantic.BaseModel): historical_summary: str = "" last_history_id: str = "" is_dirty: bool = False - last_talk: str = "" + last_talk: str = None def add_talk(self, msg: Message): msg.add_tag(MessageType.Talk.value) @@ -109,7 +109,6 @@ class BrainMemory(pydantic.BaseModel): if msg.id: if self.to_int(msg.id, 0) < self.to_int(self.last_history_id, -1): return - self.last_history_id = str(self.to_int(msg.id, 0)) self.history.append(msg.dict()) self.is_dirty = True @@ -125,3 +124,8 @@ class BrainMemory(pydantic.BaseModel): return int(v) except: return default_value + + def pop_last_talk(self): + v = self.last_talk + self.last_talk = None + return v diff --git a/metagpt/memory/longterm_memory.py b/metagpt/memory/longterm_memory.py index 041d335ac..df748037a 100644 --- a/metagpt/memory/longterm_memory.py +++ b/metagpt/memory/longterm_memory.py @@ -37,13 +37,13 @@ class LongTermMemory(Memory): self.add_batch(messages) self.msg_from_recover = False - def add(self, message: Message, **kwargs): + def add(self, message: Message): super(LongTermMemory, self).add(message) for action in self.rc.watch: if message.cause_by == action and not self.msg_from_recover: # currently, only add role's watching messages to its memory_storage # and ignore adding messages from recover repeatedly - self.memory_storage.add(message, **kwargs) + self.memory_storage.add(message) def remember(self, observed: list[Message], k=0) -> list[Message]: """ diff --git a/metagpt/memory/memory_storage.py b/metagpt/memory/memory_storage.py index 09cd67410..9afd524f0 100644 --- a/metagpt/memory/memory_storage.py +++ b/metagpt/memory/memory_storage.py @@ -5,16 +5,16 @@ @Modified By: mashenquan, 2023/8/20. Remove global configuration `CONFIG`, enable configuration support for business isolation. """ -from typing import List from pathlib import Path +from typing import List from langchain.vectorstores.faiss import FAISS from metagpt.const import DATA_PATH, MEM_TTL +from metagpt.document_store.faiss_store import FaissStore from metagpt.logs import logger from metagpt.schema import Message -from metagpt.utils.serialize import serialize_message, deserialize_message -from metagpt.document_store.faiss_store import FaissStore +from metagpt.utils.serialize import deserialize_message, serialize_message class MemoryStorage(FaissStore): @@ -37,7 +37,7 @@ class MemoryStorage(FaissStore): def recover_memory(self, role_id: str) -> List[Message]: self.role_id = role_id - self.role_mem_path = Path(DATA_PATH / f'role_mem/{self.role_id}/') + self.role_mem_path = Path(DATA_PATH / f"role_mem/{self.role_id}/") self.role_mem_path.mkdir(parents=True, exist_ok=True) self.store = self._load() @@ -54,23 +54,23 @@ class MemoryStorage(FaissStore): def _get_index_and_store_fname(self): if not self.role_mem_path: - logger.error(f'You should call {self.__class__.__name__}.recover_memory fist when using LongTermMemory') + logger.error(f"You should call {self.__class__.__name__}.recover_memory fist when using LongTermMemory") return None, None - index_fpath = Path(self.role_mem_path / f'{self.role_id}.index') - storage_fpath = Path(self.role_mem_path / f'{self.role_id}.pkl') + index_fpath = Path(self.role_mem_path / f"{self.role_id}.index") + storage_fpath = Path(self.role_mem_path / f"{self.role_id}.pkl") return index_fpath, storage_fpath def persist(self): super(MemoryStorage, self).persist() - logger.debug(f'Agent {self.role_id} persist memory into local') + logger.debug(f"Agent {self.role_id} persist memory into local") - def add(self, message: Message, **kwargs) -> bool: - """ add message into memory storage""" + def add(self, message: Message) -> bool: + """add message into memory storage""" docs = [message.content] metadatas = [{"message_ser": serialize_message(message)}] if not self.store: # init Faiss - self.store = self._write(docs, metadatas, **kwargs) + self.store = self._write(docs, metadatas) self._initialized = True else: self.store.add_texts(texts=docs, metadatas=metadatas) @@ -82,10 +82,7 @@ class MemoryStorage(FaissStore): if not self.store: return [] - resp = self.store.similarity_search_with_score( - query=message.content, - k=k - ) + resp = self.store.similarity_search_with_score(query=message.content, k=k) # filter the result which score is smaller than the threshold filtered_resp = [] for item, score in resp: diff --git a/metagpt/provider/openai_api.py b/metagpt/provider/openai_api.py index d0dd5b9d8..bf2ca7f14 100644 --- a/metagpt/provider/openai_api.py +++ b/metagpt/provider/openai_api.py @@ -226,21 +226,24 @@ class OpenAIGPTAPI(BaseGPTAPI, RateLimiter): async def get_summary(self, text: str, max_words=200, keep_language: bool = False): max_token_count = DEFAULT_MAX_TOKENS max_count = 100 + text_length = len(text) while max_count > 0: - if len(text) < max_token_count: + if text_length < max_token_count: return await self._get_summary(text=text, max_words=max_words, keep_language=keep_language) padding_size = 20 if max_token_count > 20 else 0 text_windows = self.split_texts(text, window_size=max_token_count - padding_size) + part_max_words = min(int(max_words / len(text_windows)) + 1, 100) summaries = [] for ws in text_windows: - response = await self._get_summary(text=ws, max_words=max_words, keep_language=keep_language) + response = await self._get_summary(text=ws, max_words=part_max_words, keep_language=keep_language) summaries.append(response) if len(summaries) == 1: return summaries[0] # Merged and retry text = "\n".join(summaries) + text_length = len(text) max_count -= 1 # safeguard raise openai.error.InvalidRequestError("text too long") diff --git a/metagpt/roles/assistant.py b/metagpt/roles/assistant.py index 9c80593f6..87127cbab 100644 --- a/metagpt/roles/assistant.py +++ b/metagpt/roles/assistant.py @@ -120,12 +120,12 @@ class Assistant(Role): async def refine_memory(self) -> str: history_text = self.memory.history_text - last_talk = self.memory.last_talk + last_talk = self.memory.pop_last_talk() if last_talk is None: # No user feedback, unsure if past conversation is finished. return None if history_text == "": return last_talk - history_summary = await self._llm.get_summary(history_text, max_words=500) + history_summary = await self._llm.get_summary(history_text, max_words=800, keep_language=True) await self.memory.set_history_summary( history_summary=history_summary, redis_key=CONFIG.REDIS_KEY, redis_conf=CONFIG.REDIS ) diff --git a/metagpt/utils/redis.py b/metagpt/utils/redis.py index b94eee8e2..48a18e7c9 100644 --- a/metagpt/utils/redis.py +++ b/metagpt/utils/redis.py @@ -4,6 +4,7 @@ # @Desc: { redis client } # @Date: 2022/11/28 10:12 import json +import traceback from datetime import timedelta from enum import Enum from typing import Awaitable, Callable, Dict, Optional, Union @@ -203,12 +204,19 @@ class Redis: async def get(self, key: str) -> str: if not self.is_valid() or not key: return None - v = await RedisManager.get_with_cache_info(redis_cache_info=RedisCacheInfo(key=key)) - return v + try: + v = await RedisManager.get_with_cache_info(redis_cache_info=RedisCacheInfo(key=key)) + return v + except Exception as e: + logger.exception(f"{e}, stack:{traceback.format_exc()}") + return None async def set(self, key: str, data: str, timeout_sec: int): if not self.is_valid() or not key: return - await RedisManager.set_with_cache_info( - redis_cache_info=RedisCacheInfo(key=key, timeout=timeout_sec), value=data - ) + try: + await RedisManager.set_with_cache_info( + redis_cache_info=RedisCacheInfo(key=key, timeout=timeout_sec), value=data + ) + except Exception as e: + logger.exception(f"{e}, stack:{traceback.format_exc()}") diff --git a/tests/metagpt/memory/test_longterm_memory.py b/tests/metagpt/memory/test_longterm_memory.py index 457e665fa..b77e9a955 100644 --- a/tests/metagpt/memory/test_longterm_memory.py +++ b/tests/metagpt/memory/test_longterm_memory.py @@ -4,11 +4,11 @@ @Desc : unittest of `metagpt/memory/longterm_memory.py` @Modified By: mashenquan, 2023/8/20. Remove global configuration `CONFIG`, enable configuration support for business isolation. """ -from metagpt.config import Config -from metagpt.schema import Message from metagpt.actions import BossRequirement -from metagpt.roles.role import RoleContext +from metagpt.config import Config from metagpt.memory import LongTermMemory +from metagpt.roles.role import RoleContext +from metagpt.schema import Message def test_ltm_search(): @@ -17,28 +17,28 @@ def test_ltm_search(): openai_api_key = conf.openai_api_key assert len(openai_api_key) > 20 - role_id = 'UTUserLtm(Product Manager)' - rc = RoleContext(options=conf.runtime_options, watch=[BossRequirement]) + role_id = "UTUserLtm(Product Manager)" + rc = RoleContext(watch=[BossRequirement]) ltm = LongTermMemory() ltm.recover_memory(role_id, rc) - idea = 'Write a cli snake game' - message = Message(role='BOSS', content=idea, cause_by=BossRequirement) + idea = "Write a cli snake game" + message = Message(role="BOSS", content=idea, cause_by=BossRequirement) news = ltm.remember([message]) assert len(news) == 1 - ltm.add(message, **conf.runtime_options) + ltm.add(message) - sim_idea = 'Write a game of cli snake' - sim_message = Message(role='BOSS', content=sim_idea, cause_by=BossRequirement) + sim_idea = "Write a game of cli snake" + sim_message = Message(role="BOSS", content=sim_idea, cause_by=BossRequirement) news = ltm.remember([sim_message]) assert len(news) == 0 - ltm.add(sim_message, **conf.runtime_options) + ltm.add(sim_message) - new_idea = 'Write a 2048 web game' - new_message = Message(role='BOSS', content=new_idea, cause_by=BossRequirement) + new_idea = "Write a 2048 web game" + new_message = Message(role="BOSS", content=new_idea, cause_by=BossRequirement) news = ltm.remember([new_message]) assert len(news) == 1 - ltm.add(new_message, **conf.runtime_options) + ltm.add(new_message) # restore from local index ltm_new = LongTermMemory() @@ -50,8 +50,8 @@ def test_ltm_search(): news = ltm_new.remember([sim_message]) assert len(news) == 0 - new_idea = 'Write a Battle City' - new_message = Message(role='BOSS', content=new_idea, cause_by=BossRequirement) + new_idea = "Write a Battle City" + new_message = Message(role="BOSS", content=new_idea, cause_by=BossRequirement) news = ltm_new.remember([new_message]) assert len(news) == 1