matagent/backend/retrieval_group.py

from autogen.coding import LocalCommandLineCodeExecutor
import autogen
import os
from .constant import config_list, STREAM, SILENT, WORK_DIR, CACHE
from .utils import load_agent_configs,retrieval_from_knowledge_base, web_searcher, retrieval_from_graphrag



agent_configs = load_agent_configs(os.path.join(os.path.dirname(os.path.abspath(__file__)), "config/retrieval_group.yaml"))
llm_config = {"config_list": config_list, "stream": STREAM, "cache_seed": CACHE}


def init_retrieval_group(work_dir, venv_context):
    # 向量检索
    vector_executor = LocalCommandLineCodeExecutor(
        work_dir=work_dir, 
        virtual_env_context=venv_context, 
        functions=[retrieval_from_knowledge_base]
    )

    vector_code_executor = autogen.UserProxyAgent(
        name="vector_code_executor",
        human_input_mode="NEVER",
        llm_config=False,
        code_execution_config={
            "executor": vector_executor,
        },
        description="vector_code_executor",
        max_consecutive_auto_reply=1
    )

    
    vector_searcher_name = "vector_searcher"
    vector_searcher_system_message = agent_configs[vector_searcher_name]["system_message"] + vector_executor.format_functions_for_prompt()
    vector_searcher = autogen.AssistantAgent(
        name=vector_searcher_name,
        human_input_mode="NEVER",
        system_message=vector_searcher_system_message,
        llm_config={"config_list": config_list},
        # is_termination_msg=lambda x: x.get("content", "").find("TERMINATE") >= 0,
        # human_input_mode="TERMINATE",
        description=vector_searcher_name,
        max_consecutive_auto_reply=1
    )

    # graphrag retrieval agent
    graphrag_executor = LocalCommandLineCodeExecutor(
        work_dir=work_dir, 
        virtual_env_context=venv_context, 
        functions=[retrieval_from_graphrag]
    )

    graphrag_code_executor = autogen.UserProxyAgent(
        name="graphrag_code_executor",
        human_input_mode="NEVER",
        llm_config=False,
        code_execution_config={
            "executor": graphrag_executor,
        },
        description="graphrag_code_executor",
        max_consecutive_auto_reply=1
    )

    graphrag_searcher_name = "graphrag_searcher"
    graphrag_searcher_system_message = agent_configs[graphrag_searcher_name]["system_message"] + graphrag_executor.format_functions_for_prompt()
    graphrag_searcher = autogen.AssistantAgent(
        name=graphrag_searcher_name,
        human_input_mode="NEVER",
        system_message=graphrag_searcher_system_message,
        llm_config={"config_list": config_list},
        # is_termination_msg=lambda x: x.get("content", "").find("TERMINATE") >= 0,
        # human_input_mode="TERMINATE",
        description=graphrag_searcher_name,
        max_consecutive_auto_reply=1
    )


    # web searcher with jina reader
    web_executor = LocalCommandLineCodeExecutor(
        work_dir=work_dir, 
        virtual_env_context=venv_context, 
        functions=[web_searcher]
    )

    web_code_executor = autogen.UserProxyAgent(
        name="web_code_executor",
        human_input_mode="NEVER",
        llm_config=False,
        code_execution_config={
            "executor": web_executor,
        },
        description="web_code_executor",
    )

    web_seacher_name = "web_searcher"
    web_seacher_system_message = agent_configs[web_seacher_name]["system_message"] + web_executor.format_functions_for_prompt()
    web_seacher = autogen.AssistantAgent(
        name=web_seacher_name,
        human_input_mode="NEVER",
        system_message=web_seacher_system_message,
        llm_config={"config_list": config_list},
        # is_termination_msg=lambda x: x.get("content", "").find("TERMINATE") >= 0,
        # human_input_mode="TERMINATE",
        description=web_seacher_name,
    )
    
    web_summary = autogen.AssistantAgent(
        name="web_summary",
        human_input_mode="NEVER",
        system_message="You are a web summary agent. You are given a task and a web search result. You need to summarize the web search result to answer the user's request. DONOT ANSWER THE USER'S REQUEST DIRECTLY.",
        llm_config={"config_list": config_list},
        description="web_summary",
    )

    outer_retrieval_agent = autogen.ConversableAgent(
        name="Outer_Retrieval_Admin",
        human_input_mode="NEVER",
        # human_input_mode="TERMINATE",
        code_execution_config={
            "work_dir": work_dir,
            "use_docker": False,
        },
        # is_termination_msg=lambda x: x.get("content", "").find("TERMINATE") >= 0,
        description="Outer_Retrieval_Admin",
    )

    def state_transition(last_speaker, groupchat):
        messages = groupchat.messages

        if last_speaker is outer_retrieval_agent:
            return vector_searcher
        # 向量检索
        elif last_speaker is vector_searcher:
            return vector_code_executor
        elif last_speaker is vector_code_executor:
            if "NULL" in messages[-1]["content"]:
                return 'auto'
            else:
                return outer_retrieval_agent
        
        # 知识图谱检索
        elif last_speaker is graphrag_searcher:
            return graphrag_code_executor
        elif last_speaker is graphrag_code_executor:
            return outer_retrieval_agent

        # 向量检索为空时尝试web搜索
        elif last_speaker is web_seacher:
            return web_code_executor
        elif last_speaker is web_code_executor:
            return web_summary
        elif last_speaker is web_summary:
            return outer_retrieval_agent
        

    retrieval_group = autogen.GroupChat(
        agents=[outer_retrieval_agent, vector_searcher, vector_code_executor, graphrag_searcher, graphrag_code_executor, web_seacher, web_code_executor, web_summary],
        messages=[],
        speaker_selection_method=state_transition,  # With two agents, this is equivalent to a 1:1 conversation.
        allow_repeat_speaker=False,
        max_round=8,
    )

    inner_retrieval_admin = autogen.GroupChatManager(
        name="Inner_Retrieval_Admin",
        groupchat=retrieval_group,
        # is_termination_msg=lambda x: x.get("content", "").find("TERMINATE") >= 0,
        llm_config={"config_list": config_list},
        human_input_mode="NEVER",
        description="Inner_Retrieval_Admin",
    )

    return inner_retrieval_admin, outer_retrieval_agent