如何让LangChain Agent按正确顺序选择对应工具？

核心问题分析 #

原代码的三个工具都绑定了同一个qa.run（内部检索链），完全不符合功能需求；同时工具描述模糊，Agent无法准确判断触发时机，导致工具选择逻辑失效。

具体调整步骤 #

1. 拆分工具执行逻辑 #

• 内部知识库工具：保留原RetrievalQA检索链，新增检索结果有效性判断
• 外部知识库工具：直接调用大模型原生能力，无需访问内部库
• 日常对话工具：调用大模型并注入友好角色设定，生成自然回应

2. 优化工具描述 #

给每个工具明确触发条件与优先级，帮助Agent精准判断调用时机

3. 调整RetrievalQA配置 #

开启return_source_documents=True，让Agent能识别内部库是否检索到有效内容

4. 优化对话记忆 #

调高k值保留对话上下文，让日常对话更连贯自然

完整修正代码 #

from langchain.agents import Tool
from langchain.chat_models import ChatOpenAI
from langchain.chains.conversation.memory import ConversationBufferWindowMemory
from langchain.chains import RetrievalQA
from langchain.agents import initialize_agent
from chroma_database import ChromaDatabase
from langchain.embeddings import OpenAIEmbeddings
from parameters import EMBEDDING_MODEL, BUCKET_NAME, COLLECTION_NAME

embeddings = OpenAIEmbeddings(model=EMBEDDING_MODEL)
chroma = ChromaDatabase(embedding_function=embeddings, 
                    persist_directory='database/vectors/', 
                    bucket_name=BUCKET_NAME,
                    collection_name=COLLECTION_NAME)

# 基础大模型实例
llm = ChatOpenAI(
    model_name='gpt-3.5-turbo',
    temperature=0.0
)

# 对话记忆：保留最近2轮对话，维持上下文连贯性
conversational_memory = ConversationBufferWindowMemory(
    memory_key='chat_history',
    k=2,
    return_messages=True
)

# 内部知识库检索链：开启源文档返回，用于判断检索有效性
qa_internal = RetrievalQA.from_chain_type(
    llm=llm,
    chain_type="stuff",
    retriever=chroma.db.as_retriever(),
    return_source_documents=True  # 关键配置：返回检索到的原始文档
)

# 封装内部知识库工具逻辑：判断检索结果并返回对应内容
def run_internal_knowledge(query):
    result = qa_internal({"query": query})
    # 无有效检索结果时，返回明确提示引导Agent切换工具
    if not result['source_documents']:
        return "No relevant information found in internal knowledge base. Please use external knowledge base."
    return result['result']

# 外部知识库工具逻辑：直接调用大模型原生能力
def run_external_knowledge(query):
    return llm.predict(query)

# 日常对话工具逻辑：注入友好角色设定
def run_natural_conversation(query):
    prompt = f"Act as a friendly chat partner, respond naturally to: {query}"
    return llm.predict(prompt)

# 定义工具列表，每个工具的描述明确触发规则与优先级
tools = [
    Tool(
        name='Knowledge Internal Base',
        func=run_internal_knowledge,
        description=(
            'Always use this tool FIRST when answering questions related to internal stored knowledge (e.g., Pepito-related content). '
            'If it returns "No relevant information found", switch to the external knowledge tool.'
        )
    ),
    Tool(
        name='Knowledge External Base',
        func=run_external_knowledge,
        description=(
            'Use this tool ONLY when the internal knowledge base returns no relevant information. '
            'For public information not stored internally (e.g., Tom Cruise movie details).'
        )
    ),
    Tool(
        name='Natural Conversation',
        func=run_natural_conversation,
        description=(
            'Use this tool ONLY for casual daily conversations (e.g., greetings, "how are you?", small talk). '
            'Respond in a warm, human-like tone.'
        )
    )
]

# 初始化Agent：适配对话场景的agent类型
agent = initialize_agent(
    agent='chat-conversational-react-description',
    tools=tools,
    llm=llm,
    verbose=True,
    max_iterations=3,
    early_stopping_method='generate',
    memory=conversational_memory
)

# 测试用例
agent.run("What Pepito said?")
agent.run("What Tom Cruise said in the movie Impossible Mission 1?")
agent.run("Hello, how are you?")

关键调整说明 #

1. 工具逻辑解耦：每个工具绑定独立的执行函数，彻底解决原代码中工具功能混淆的问题
2. 检索结果判断：通过源文档返回状态，让Agent能明确感知内部库是否有答案，实现工具切换逻辑
3. 工具描述强化：明确了调用优先级与触发边界，比如内部工具必须优先使用，日常对话工具仅用于闲聊场景
4. 对话记忆优化：将k值从0调整为2，保留对话上下文，让日常对话更符合自然交流逻辑

内容的提问来源于stack exchange，提问作者Eric Bellet