GraphRAG × AutoGen × Ollama × Chainlit = 本地免费多代理 RAG 超级 AI 助手

# Mistral for GraphRAG Inferenceollama pull mistral# Nomic-Embed-Text for GraphRAG Embeddingollama pull nomic-embed-text# LLama3 for Autogen Inferenceollama pull llama3# Host Ollama on a local server: http://localhost:11434ollama serve

# Create and activate a conda environmentconda create -n RAG_agents python=3.12conda activate RAG_agents# Lite-LLM proxy server for Ollamapip install 'litellm[proxy]'# Install Ollamapip install ollama# Microsoft AutoGenpip install pyautogen "pyautogen[retrievechat]" # Microsoft GraphRAGpip install graphrag# Text-Token Encoder-Decoderpip install tiktoken# Chainlit Python applicationpip install chainlit# Clone my Git-hub repositorygit clone https://github.com/karthik-codex/autogen_graphRAG.git# (BONUS) To Convert PDF files to Markdown for GraphRAG pip install marker-pdf# (BONUS) Only if you installed Marker-pdf since it removes GPU CUDA support by defaultconda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

#make a new folder "input" to place your input files for GraphRAG (.txt or .md)mkdir -p ./input# Initialize GraphRAG to create the required files and folders in the root dirpython -m graphrag.index --init  --root .# Move the settings.yaml file to replace the one created by GraphRAG --initmv ./utils/settings.yaml ./

encoding_model: cl100k_baseskip_workflows: []llm:  api_key: ${GRAPHRAG_API_KEY}  type: openai_chat # or azure_openai_chat  model: mistral  model_supports_json: true  api_base: http://localhost:11434/v1 
embeddings:  async_mode: threaded # or asyncio  llm:    api_key: ${GRAPHRAG_API_KEY}    type: openai_embedding # or azure_openai_embedding    model: nomic_embed_text    api_base: http://localhost:11434/api 
input:  #Change input file pattern to.md, or .txt  type: file # or blob  file_type: text # or csv  base_dir: "input"  file_encoding: utf-8  file_pattern: ".*\\.md$"

# Create knowledge graph - this takes some timepython -m graphrag.index --root .# Test GraphRAGpython -m graphrag.query --root . --method global "<insert your query>"

# start server from terminallitellm --model ollama_chat/llama3# run app from another terminalchainlit run appUI.py

import autogenfrom rich import printimport chainlit as clfrom typing_extensions import Annotatedfrom chainlit.input_widget import (   Select, Slider, Switch)from autogen import AssistantAgent, UserProxyAgentfrom utils.chainlit_agents import ChainlitUserProxyAgent, ChainlitAssistantAgentfrom graphrag.query.cli import run_global_search, run_local_search

@cl.on_chat_startasync def on_chat_start():  try:    settings = await cl.ChatSettings(            [                      Switch(id="Search_type", label="(GraphRAG) Local Search", initial=True),                       Select(                    id="Gen_type",                    label="(GraphRAG) Content Type",                    values=["prioritized list", "single paragraph", "multiple paragraphs", "multiple-page report"],                    initial_index=1,                ),                          Slider(                    id="Community",                    label="(GraphRAG) Community Level",                    initial=0,                    min=0,                    max=2,                    step=1,                ),            ]        ).send()    response_type = settings["Gen_type"]    community = settings["Community"]    local_search = settings["Search_type"]    cl.user_session.set("Gen_type", response_type)    cl.user_session.set("Community", community)    cl.user_session.set("Search_type", local_search)    retriever   = AssistantAgent(       name="Retriever",        llm_config=llm_config_autogen,        system_message="""Only execute the function query_graphRAG to look for context.                     Output 'TERMINATE' when an answer has been provided.""",       max_consecutive_auto_reply=1,       human_input_mode="NEVER",        description="Retriever Agent"    )    user_proxy = ChainlitUserProxyAgent(        name="User_Proxy",        human_input_mode="ALWAYS",        llm_config=llm_config_autogen,        is_termination_msg=lambda x: x.get("content", "").rstrip().endswith("TERMINATE"),        code_execution_config=False,        system_message='''A human admin. Interact with the retriever to provide any context''',        description="User Proxy Agent"    )    print("Set agents.")    cl.user_session.set("Query Agent", user_proxy)    cl.user_session.set("Retriever", retriever)    msg = cl.Message(content=f"""Hello! What task would you like to get done today?                           """,                      author="User_Proxy")    await msg.send()    print("Message sent.")      except Exception as e:    print("Error: ", e)    pass

@cl.on_settings_updateasync def setup_agent(settings):    response_type = settings["Gen_type"]    community = settings["Community"]    local_search = settings["Search_type"]    cl.user_session.set("Gen_type", response_type)    cl.user_session.set("Community", community)    cl.user_session.set("Search_type", local_search)    print("on_settings_update", settings)

@cl.on_messageasync def run_conversation(message: cl.Message):    print("Running conversation")    CONTEXT = message.content    MAX_ITER = 10    INPUT_DIR = None    ROOT_DIR = '.'    RESPONSE_TYPE = cl.user_session.get("Gen_type")    COMMUNITY = cl.user_session.get("Community")    LOCAL_SEARCH = cl.user_session.get("Search_type")    print("Setting groupchat")    retriever   = cl.user_session.get("Retriever")    user_proxy  = cl.user_session.get("Query Agent")    def state_transition(last_speaker, groupchat):        messages = groupchat.messages        if last_speaker is user_proxy:            return retriever        if last_speaker is retriever:            if messages[-1]["content"].lower() not in ['math_expert','physics_expert']:                return user_proxy            else:                if messages[-1]["content"].lower() == 'math_expert':                    return user_proxy                else:                    return user_proxy        else:            pass            return None    async def query_graphRAG(          question: Annotated[str, 'Query string containing information that you want from RAG search']                          ) -> str:        if LOCAL_SEARCH:            result = run_local_search(INPUT_DIR, ROOT_DIR, COMMUNITY ,RESPONSE_TYPE, question)        else:            result = run_global_search(INPUT_DIR, ROOT_DIR, COMMUNITY ,RESPONSE_TYPE, question)        await cl.Message(content=result).send()        return result    for caller in [retriever]:        d_retrieve_content = caller.register_for_llm(            description="retrieve content for code generation and question answering.", api_style="function"        )(query_graphRAG)    for agents in [user_proxy, retriever]:        agents.register_for_execution()(d_retrieve_content)    groupchat = autogen.GroupChat(        agents=[user_proxy, retriever],        messages=[],        max_round=MAX_ITER,        speaker_selection_method=state_transition,        allow_repeat_speaker=True,    )    manager = autogen.GroupChatManager(groupchat=groupchat,                                       llm_config=llm_config_autogen,                                        is_termination_msg=lambda x: x.get("content", "") and x.get("content", "").rstrip().endswith("TERMINATE"),                                       code_execution_config=False,                                       )    # -------------------- Conversation Logic. Edit to change your first message based on the Task you want to get done. ----------------------------- #     if len(groupchat.messages) == 0:       await cl.make_async(user_proxy.initiate_chat)( manager, message=CONTEXT, )    elif len(groupchat.messages) < MAX_ITER:      await cl.make_async(user_proxy.send)( manager, message=CONTEXT, )    elif len(groupchat.messages) == MAX_ITER:        await cl.make_async(user_proxy.send)( manager, message="exit", )