Haystack 2.12非Agent RAG Pipeline添加DuckDuckGo搜索工具失败求助

核心问题拆解 #

• LLM无工具调用意识：当前Prompt完全未告知模型存在WebSearch工具，导致模型无法判断何时需要触发搜索，只会返回预设的无实时数据提示。
• 工具调用流程断裂：即便触发工具调用，搜索结果也未回传给LLM，无法基于新数据生成回答。
• LLM未配置工具支持：Haystack 2.x中，OpenAIGenerator需显式传入tools参数，才能让模型生成符合格式的工具调用指令。
• Pipeline路由不完整：缺少工具调用结果到LLM的回流路径，也未明确最终回复的输出端口。

具体修复步骤 #

1. 修改Prompt模板，添加工具调用规则 #

在Prompt中明确告知模型工具的存在、适用场景和调用格式：

def _create_system_prompt(self):
    return """
    You are an AI assistant designed to provide **clear, concise, and accurate** responses. Your goal is to **help users efficiently** while providing recommendations only if the question relates to something you can recommend. Your responses should be **direct, informative, and polite**, without unnecessary details or filler content.

    **可用工具**:
    当你需要实时数据或当前文档中没有的信息时，必须调用WebSearch工具，调用格式严格遵循：
    <|tool_call_begin|>[{"name": "WebSearch", "parameters": {"query": "你的搜索关键词"}}]<|tool_call_end|>

    **Conversation Context**:
    {% if conversation_history %}
        Previous conversation history:
        {{ conversation_history }}
    {% else %}
        This is a new conversation.
    {% endif %}

    **Documents**:
    {% for doc in documents %}
        {{ doc.content }}
    {% endfor %}

    **Image Description**
    {% if image_description %}
        An image description has been provided. Use the description to assist with the response, including any user-related details that might be inferred.
        {{ image_description }}
    {% else %}
        No image description provided.
    {% endif %}

    **Question**: {{ question }}

    **Answer**:
    """

2. 配置OpenAIGenerator支持工具调用 #

初始化LLM时传入tools参数，让模型知晓可用工具：

def _initialize_language_model(self):
    api_key = "API_KEY"
    model_name = self.model
    # 传入WebTool，启用模型的工具调用能力
    return OpenAIGenerator(api_key=Secret.from_token(api_key), model=model_name, tools=[self.web_tool])

3. 完善Pipeline的路由与连接 #

补充工具调用结果到LLM的回流路径，并设置输出端口：

def _initialize_rag_pipeline(self):
    pipeline = Pipeline()
    pipeline.add_component("retriever", self.retriever)
    pipeline.add_component("prompt_builder", self.prompt_builder)
    pipeline.add_component("llm", self.llm)
    pipeline.add_component("router", ConditionalRouter(self.web_route))
    pipeline.add_component("tool_invoker", ToolInvoker(tools=[self.web_tool]))

    # 原有连接保留
    pipeline.connect("retriever.documents", "prompt_builder.documents")
    pipeline.connect("prompt_builder", "llm.prompt")
    pipeline.connect("llm.replies", "router.replies")
    pipeline.connect("router.there_are_tool_calls", "tool_invoker.messages")
    
    # 新增：工具调用结果回传给LLM，生成最终回答
    pipeline.connect("tool_invoker.replies", "llm.messages")
    
    # 设置输出端口，覆盖两种回复场景
    pipeline.set_outputs(["router.final_replies", "llm.replies"])
    return pipeline

4. 添加Pipeline执行与结果处理方法 #

新增run方法处理输入输出，判断返回直接回复或工具调用后的结果：

def run(self, question: str, conversation_history: List[ChatMessage] = None, image_description: str = None):
    inputs = {
        "retriever": {"query": question},
        "prompt_builder": {
            "question": question,
            "conversation_history": conversation_history or [],
            "image_description": image_description or ""
        }
    }
    result = self.rag_pipeline.run(inputs)
    
    # 优先返回无需工具调用的直接回复
    if "final_replies" in result and result["final_replies"]:
        return result["final_replies"][0].content
    # 其次返回工具调用后的生成回复
    elif "replies" in result and result["replies"]:
        return result["replies"][0].content
    # 兜底返回无数据提示
    else:
        return "I didn't have real-time data"

5. 调整类初始化顺序 #

由于LLM需要依赖WebTool，需提前初始化web_tool：

def __init__(self, model="gpt-4o-mini"):
    self.model = model
    self.prompt_template = self._create_system_prompt()
    self.document_store = self._initialize_document_store()
    self.web_tool = self._initialize_web_tool()  # 提前初始化，供LLM使用
    self.llm = self._initialize_language_model()
    self.retriever = self._initialize_retriever()
    self.prompt_builder = self._initialize_prompt_builder()
    self.web_route = self._initialize_web_route()
    self.rag_pipeline = self._initialize_rag_pipeline()

内容的提问来源于stack exchange，提问作者Abstract