feat(agent):Agent supports conversation context (#2230)

Co-authored-by: Fangyin Cheng <staneyffer@gmail.com>

.env.template

@@ -333,4 +333,9 @@ DBGPT_LOG_LEVEL=INFO
 # FIN_REPORT_MODEL=/app/models/bge-large-zh
 
 ## Turn off notebook display Python flow , which is enabled by default
-NOTE_BOOK_ENABLE=False
+NOTE_BOOK_ENABLE=False
+
+## The agent historical message retention configuration defaults to the last two rounds.
+# MESSAGES_KEEP_START_ROUNDS=0
+# MESSAGES_KEEP_END_ROUNDS=2
+

dbgpt/_private/config.py

@@ -361,6 +361,13 @@ class Config(metaclass=Singleton):
         )
         self.NOTE_BOOK_ROOT: str = os.getenv("NOTE_BOOK_ROOT", os.path.expanduser("~"))
 
+        self.MESSAGES_KEEP_START_ROUNDS: int = int(
+            os.getenv("MESSAGES_KEEP_START_ROUNDS", 0)
+        )
+        self.MESSAGES_KEEP_END_ROUNDS: int = int(
+            os.getenv("MESSAGES_KEEP_END_ROUNDS", 2)
+        )
+
     @property
     def local_db_manager(self) -> "ConnectorManager":
         from dbgpt.datasource.manages import ConnectorManager

dbgpt/agent/core/agent.py

@@ -27,6 +27,8 @@ class Agent(ABC):
         silent: Optional[bool] = False,
         is_retry_chat: bool = False,
         last_speaker_name: Optional[str] = None,
+        rely_messages: Optional[List[AgentMessage]] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
     ) -> None:
         """Send a message to recipient agent.
 
@@ -52,6 +54,8 @@ class Agent(ABC):
         is_recovery: Optional[bool] = False,
         is_retry_chat: bool = False,
         last_speaker_name: Optional[str] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
+        rely_messages: Optional[List[AgentMessage]] = None,
     ) -> None:
         """Receive a message from another agent.
 
@@ -74,6 +78,7 @@ class Agent(ABC):
         sender: Agent,
         reviewer: Optional[Agent] = None,
         rely_messages: Optional[List[AgentMessage]] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
         is_retry_chat: bool = False,
         last_speaker_name: Optional[str] = None,
         **kwargs,

dbgpt/agent/core/base_agent.py

@@ -209,6 +209,8 @@ class ConversableAgent(Role, Agent):
         silent: Optional[bool] = False,
         is_retry_chat: bool = False,
         last_speaker_name: Optional[str] = None,
+        rely_messages: Optional[List[AgentMessage]] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
     ) -> None:
         """Send a message to recipient agent."""
         with root_tracer.start_span(
@@ -232,6 +234,8 @@ class ConversableAgent(Role, Agent):
                 silent=silent,
                 is_retry_chat=is_retry_chat,
                 last_speaker_name=last_speaker_name,
+                historical_dialogues=historical_dialogues,
+                rely_messages=rely_messages,
             )
 
     async def receive(
@@ -244,6 +248,8 @@ class ConversableAgent(Role, Agent):
         is_recovery: Optional[bool] = False,
         is_retry_chat: bool = False,
         last_speaker_name: Optional[str] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
+        rely_messages: Optional[List[AgentMessage]] = None,
     ) -> None:
         """Receive a message from another agent."""
         with root_tracer.start_span(
@@ -272,6 +278,8 @@ class ConversableAgent(Role, Agent):
                         reviewer=reviewer,
                         is_retry_chat=is_retry_chat,
                         last_speaker_name=last_speaker_name,
+                        historical_dialogues=historical_dialogues,
+                        rely_messages=rely_messages,
                     )
                 else:
                     reply = await self.generate_reply(
@@ -279,6 +287,8 @@ class ConversableAgent(Role, Agent):
                         sender=sender,
                         reviewer=reviewer,
                         is_retry_chat=is_retry_chat,
+                        historical_dialogues=historical_dialogues,
+                        rely_messages=rely_messages,
                     )
 
                 if reply is not None:
@@ -289,6 +299,7 @@ class ConversableAgent(Role, Agent):
         received_message: Optional[AgentMessage],
         sender: Agent,
         rely_messages: Optional[List[AgentMessage]] = None,
+        **kwargs,
     ) -> Dict[str, Any]:
         """Prepare the parameters for the act method."""
         return {}
@@ -300,6 +311,7 @@ class ConversableAgent(Role, Agent):
         sender: Agent,
         reviewer: Optional[Agent] = None,
         rely_messages: Optional[List[AgentMessage]] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
         is_retry_chat: bool = False,
         last_speaker_name: Optional[str] = None,
         **kwargs,
@@ -361,9 +373,10 @@ class ConversableAgent(Role, Agent):
                     f"Depends on the number of historical messages:{len(rely_messages) if rely_messages else 0}！"  # noqa
                 )
                 thinking_messages, resource_info = await self._load_thinking_messages(
-                    received_message,
-                    sender,
-                    rely_messages,
+                    received_message=received_message,
+                    sender=sender,
+                    rely_messages=rely_messages,
+                    historical_dialogues=historical_dialogues,
                     context=reply_message.get_dict_context(),
                     is_retry_chat=is_retry_chat,
                 )
@@ -400,7 +413,10 @@ class ConversableAgent(Role, Agent):
                     span.metadata["comments"] = comments
 
                 act_extent_param = self.prepare_act_param(
-                    received_message, sender, rely_messages
+                    received_message=received_message,
+                    sender=sender,
+                    rely_messages=rely_messages,
+                    historical_dialogues=historical_dialogues,
                 )
                 with root_tracer.start_span(
                     "agent.generate_reply.act",
@@ -620,6 +636,8 @@ class ConversableAgent(Role, Agent):
         is_retry_chat: bool = False,
         last_speaker_name: Optional[str] = None,
         message_rounds: int = 0,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
+        rely_messages: Optional[List[AgentMessage]] = None,
         **context,
     ):
         """Initiate a chat with another agent.
@@ -652,6 +670,8 @@ class ConversableAgent(Role, Agent):
                 agent_message,
                 recipient,
                 reviewer,
+                historical_dialogues=historical_dialogues,
+                rely_messages=rely_messages,
                 request_reply=request_reply,
                 is_retry_chat=is_retry_chat,
                 last_speaker_name=last_speaker_name,
@@ -825,6 +845,38 @@ class ConversableAgent(Role, Agent):
 
         return can_uses
 
+    def convert_to_agent_message(
+        self,
+        gpts_messages: List[GptsMessage],
+        is_rery_chat: bool = False,
+    ) -> Optional[List[AgentMessage]]:
+        """Convert gptmessage to agent message."""
+        oai_messages: List[AgentMessage] = []
+        # Based on the current agent, all messages received are user, and all messages
+        # sent are assistant.
+        if not gpts_messages:
+            return None
+        for item in gpts_messages:
+            # Message conversion, priority is given to converting execution results,
+            # and only model output results will be used if not.
+            content = item.content
+            oai_messages.append(
+                AgentMessage(
+                    content=content,
+                    context=(
+                        json.loads(item.context) if item.context is not None else None
+                    ),
+                    action_report=ActionOutput.from_dict(json.loads(item.action_report))
+                    if item.action_report
+                    else None,
+                    name=item.sender,
+                    rounds=item.rounds,
+                    model_name=item.model_name,
+                    success=item.is_success,
+                )
+            )
+        return oai_messages
+
     async def _a_select_llm_model(
         self, excluded_models: Optional[List[str]] = None
     ) -> str:
@@ -959,6 +1011,7 @@ class ConversableAgent(Role, Agent):
         received_message: AgentMessage,
         sender: Agent,
         rely_messages: Optional[List[AgentMessage]] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
         context: Optional[Dict[str, Any]] = None,
         is_retry_chat: bool = False,
     ) -> Tuple[List[AgentMessage], Optional[Dict]]:
@@ -1020,13 +1073,27 @@ class ConversableAgent(Role, Agent):
                     role=ModelMessageRoleType.SYSTEM,
                 )
             )
-        if user_prompt:
-            agent_messages.append(
-                AgentMessage(
-                    content=user_prompt,
-                    role=ModelMessageRoleType.HUMAN,
-                )
+        # 关联上下文的历史消息
+        if historical_dialogues:
+            for i in range(len(historical_dialogues)):
+                if i % 2 == 0:
+                    # 偶数开始， 偶数是用户信息
+                    message = historical_dialogues[i]
+                    message.role = ModelMessageRoleType.HUMAN
+                    agent_messages.append(message)
+                else:
+                    # 奇数是AI信息
+                    message = historical_dialogues[i]
+                    message.role = ModelMessageRoleType.AI
+                    agent_messages.append(message)
+
+        # 当前的用户输入信息
+        agent_messages.append(
+            AgentMessage(
+                content=user_prompt,
+                role=ModelMessageRoleType.HUMAN,
             )
+        )
 
         return agent_messages, resource_references
 

dbgpt/agent/core/base_team.py

@@ -161,8 +161,9 @@ class ManagerAgent(ConversableAgent, Team):
         received_message: AgentMessage,
         sender: Agent,
         rely_messages: Optional[List[AgentMessage]] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
         context: Optional[Dict[str, Any]] = None,
-        is_retry_chat: Optional[bool] = False,
+        is_retry_chat: bool = False,
     ) -> Tuple[List[AgentMessage], Optional[Dict]]:
         """Load messages for thinking."""
         return [AgentMessage(content=received_message.content)], None

dbgpt/agent/core/memory/gpts/base.py

@@ -58,7 +58,7 @@ class GptsMessage:
     receiver: str
     role: str
     content: str
-    rounds: Optional[int]
+    rounds: int = 0
     is_success: bool = True
     app_code: Optional[str] = None
     app_name: Optional[str] = None

dbgpt/agent/core/memory/gpts/gpts_memory.py

@@ -127,8 +127,11 @@ class GptsMemory:
         await self.push_message(conv_id)
 
     async def get_messages(self, conv_id: str) -> List[GptsMessage]:
-        """Get conversation message."""
-        return self.messages_cache[conv_id]
+        """Get message by conv_id."""
+        messages = self.messages_cache[conv_id]
+        if not messages:
+            messages = self.message_memory.get_by_conv_id(conv_id)
+        return messages
 
     async def get_agent_messages(
         self, conv_id: str, agent_role: str

dbgpt/agent/core/plan/awel/team_awel_layout.py

@@ -2,7 +2,7 @@
 
 import logging
 from abc import ABC, abstractmethod
-from typing import Optional, cast
+from typing import List, Optional, cast
 
 from dbgpt._private.config import Config
 from dbgpt._private.pydantic import (
@@ -114,6 +114,8 @@ class AWELBaseManager(ManagerAgent, ABC):
         is_recovery: Optional[bool] = False,
         is_retry_chat: bool = False,
         last_speaker_name: Optional[str] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
+        rely_messages: Optional[List[AgentMessage]] = None,
     ) -> None:
         """Recive message by base team."""
         if request_reply is False or request_reply is None:

dbgpt/agent/core/plan/planner_agent.py

@@ -169,6 +169,7 @@ assistants:[
         received_message: Optional[AgentMessage],
         sender: Agent,
         rely_messages: Optional[List[AgentMessage]] = None,
+        **kwargs,
     ) -> Dict[str, Any]:
         """Prepare the parameters for the act method."""
         return {

dbgpt/agent/core/user_proxy_agent.py

@@ -1,5 +1,5 @@
 """A proxy agent for the user."""
-from typing import Optional
+from typing import List, Optional
 
 from .. import ActionOutput, Agent, AgentMessage
 from .base_agent import ConversableAgent
@@ -39,6 +39,8 @@ class UserProxyAgent(ConversableAgent):
         is_recovery: Optional[bool] = False,
         is_retry_chat: bool = False,
         last_speaker_name: Optional[str] = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
+        rely_messages: Optional[List[AgentMessage]] = None,
     ) -> None:
         """Receive a message from another agent."""
         if not silent:

dbgpt/serve/agent/agents/controller.py

@@ -206,11 +206,44 @@ class MultiAgents(BaseComponent, ABC):
                 if not gpt_app:
                     raise ValueError(f"Not found app {gpts_name}!")
 
+        historical_dialogues: List[GptsMessage] = []
         if not is_retry_chat:
-            # 新建gpts对话记录
+            # Create a new gpts conversation record
             gpt_app: GptsApp = self.gpts_app.app_detail(gpts_name)
             if not gpt_app:
                 raise ValueError(f"Not found app {gpts_name}!")
+
+            ## When creating a new gpts conversation record, determine whether to include the history of previous topics according to the application definition.
+            ## TODO BEGIN
+            # Temporarily use system configuration management, and subsequently use application configuration management
+            if CFG.MESSAGES_KEEP_START_ROUNDS and CFG.MESSAGES_KEEP_START_ROUNDS > 0:
+                gpt_app.keep_start_rounds = CFG.MESSAGES_KEEP_START_ROUNDS
+            if CFG.MESSAGES_KEEP_END_ROUNDS and CFG.MESSAGES_KEEP_END_ROUNDS > 0:
+                gpt_app.keep_end_rounds = CFG.MESSAGES_KEEP_END_ROUNDS
+            ## TODO END
+
+            if gpt_app.keep_start_rounds > 0 or gpt_app.keep_end_rounds > 0:
+                if gpts_conversations and len(gpts_conversations) > 0:
+                    rely_conversations = []
+                    if gpt_app.keep_start_rounds + gpt_app.keep_end_rounds < len(
+                        gpts_conversations
+                    ):
+                        if gpt_app.keep_start_rounds > 0:
+                            front = gpts_conversations[gpt_app.keep_start_rounds :]
+                            rely_conversations.extend(front)
+                        if gpt_app.keep_end_rounds > 0:
+                            back = gpts_conversations[-gpt_app.keep_end_rounds :]
+                            rely_conversations.extend(back)
+                    else:
+                        rely_conversations = gpts_conversations
+                    for gpts_conversation in rely_conversations:
+                        temps: List[GptsMessage] = await self.memory.get_messages(
+                            gpts_conversation.conv_id
+                        )
+                        if temps and len(temps) > 1:
+                            historical_dialogues.append(temps[0])
+                            historical_dialogues.append(temps[-1])
+
             self.gpts_conversations.add(
                 GptsConversationsEntity(
                     conv_id=agent_conv_id,
@@ -277,6 +310,8 @@ class MultiAgents(BaseComponent, ABC):
                         is_retry_chat,
                         last_speaker_name=last_speaker_name,
                         init_message_rounds=message_round,
+                        enable_verbose=enable_verbose,
+                        historical_dialogues=historical_dialogues,
                         **ext_info,
                     )
                 )
@@ -418,6 +453,8 @@ class MultiAgents(BaseComponent, ABC):
         link_sender: ConversableAgent = None,
         app_link_start: bool = False,
         enable_verbose: bool = True,
+        historical_dialogues: Optional[List[GptsMessage]] = None,
+        rely_messages: Optional[List[GptsMessage]] = None,
         **ext_info,
     ):
         gpts_status = Status.COMPLETE.value
@@ -529,6 +566,10 @@ class MultiAgents(BaseComponent, ABC):
                     is_retry_chat=is_retry_chat,
                     last_speaker_name=last_speaker_name,
                     message_rounds=init_message_rounds,
+                    historical_dialogues=user_proxy.convert_to_agent_message(
+                        historical_dialogues
+                    ),
+                    rely_messages=rely_messages,
                     **ext_info,
                 )
 

dbgpt/serve/agent/agents/expand/app_start_assisant_agent.py

@@ -93,6 +93,8 @@ class StartAppAssistantAgent(ConversableAgent):
         is_recovery: Optional[bool] = False,
         is_retry_chat: bool = False,
         last_speaker_name: str = None,
+        historical_dialogues: Optional[List[AgentMessage]] = None,
+        rely_messages: Optional[List[AgentMessage]] = None,
     ) -> None:
         await self._a_process_received_message(message, sender)
         if request_reply is False or request_reply is None:

dbgpt/serve/agent/db/gpts_app.py

@@ -135,6 +135,10 @@ class GptsApp(BaseModel):
     recommend_questions: Optional[List[RecommendQuestion]] = []
     admins: List[str] = Field(default_factory=list)
 
+    # By default, keep the last two rounds of conversation records as the context
+    keep_start_rounds: int = 0
+    keep_end_rounds: int = 0
+
     def to_dict(self):
         return {k: self._serialize(v) for k, v in self.__dict__.items()}
 
@@ -170,6 +174,8 @@ class GptsApp(BaseModel):
             owner_avatar_url=d.get("owner_avatar_url", None),
             recommend_questions=d.get("recommend_questions", []),
             admins=d.get("admins", []),
+            keep_start_rounds=d.get("keep_start_rounds", 0),
+            keep_end_rounds=d.get("keep_end_rounds", 2),
         )
 
     @model_validator(mode="before")
@@ -547,6 +553,8 @@ class GptsAppDao(BaseDao):
                         "published": app_info.published,
                         "details": [],
                         "admins": [],
+                        # "keep_start_rounds": app_info.keep_start_rounds,
+                        # "keep_end_rounds": app_info.keep_end_rounds,
                     }
                 )
                 for app_info in app_entities
@@ -918,6 +926,8 @@ class GptsAppDao(BaseDao):
             app_entity.icon = gpts_app.icon
             app_entity.team_context = _parse_team_context(gpts_app.team_context)
             app_entity.param_need = json.dumps(gpts_app.param_need)
+            app_entity.keep_start_rounds = gpts_app.keep_start_rounds
+            app_entity.keep_end_rounds = gpts_app.keep_end_rounds
             session.merge(app_entity)
 
             old_details = session.query(GptsAppDetailEntity).filter(