自定义LangChain工具无法完成Agent流水线，Discord功能异常求助

开发Discord消息AI助手时，自定义Discord工具函数返回的成功提示（如Successfully sent Can you grab some apples on your way home? to John）未被CHAT_CONVERSATIONAL_REACT_DESCRIPTION类型的Agent识别为有效工具响应，导致Agent错误输出"There was no tool response"。

核心原因 #

1. 内部LLM的JSON格式要求错误：原Prompt要求输出的JSON缺少外层大括号，虽然本次运行侥幸解析成功，但存在格式合法性隐患，易导致json.loads失败。
2. 纯自然语言响应干扰Agent解析：该类型Agent对工具响应格式敏感，无标识的纯自然语言字符串可能被误判为对话内容，而非工具执行结果。
3. 返回字符串末尾存在多余空格：可能干扰Agent对响应边界的识别。

解决方案 #

1. 修复内部LLM的JSON输出格式 #

修改Discord工具内部的PromptTemplate，确保生成合法的JSON对象（用双大括号{{}}转义避免模板解析冲突）：

prompt = PromptTemplate(
    template="""
        You are a message interpreter for Discord messages. Extract the recipient and core message content from: {message}.
        Output ONLY a valid JSON object with no extra text, following this structure:
        {{
          "recipient": "name_of_recipient",
          "message": "core_message_text"
        }}
    """, 
    input_variables=["message"],
)

2. 给工具响应添加明确标识 #

返回带特殊标记的字符串，帮助Agent快速识别工具执行结果：

def discord(message):
    # ... 保留其他代码 ...
    try:
        jsonResp = json.loads(response)
        # 增加明确标识，避免与普通对话混淆
        return f"[Discord Tool Success] Sent '{jsonResp['message']}' to {jsonResp['recipient']}"
    except json.JSONDecodeError:
        return "[Discord Tool Error] Failed to parse message content"

3. 简化工具定义 #

去掉不必要的lambda包装，直接引用函数：

tools = [
    # Fibonacci工具不变
    Tool(
        name="Discord",
        func=discord,  # 直接使用函数，无需lambda
        description="Use when you need to send a message via Discord. Input should include recipient and message content in natural language."
    ),
]

4. 可选：使用StructuredTool明确输入输出结构 #

通过StructuredTool定义工具的输入输出规范，帮助Agent更清晰理解工具用法：

from langchain.tools import StructuredTool
from pydantic import BaseModel, Field

class DiscordInput(BaseModel):
    message: str = Field(description="Natural language instruction containing recipient and message content for Discord")

discord_tool = StructuredTool.from_function(
    func=discord,
    name="Discord",
    description="Use when you need to send a message via Discord. Input should include recipient and message content in natural language.",
    args_schema=DiscordInput,
)

tools = [fib_tool, discord_tool]

修改后的完整代码示例 #

import json
from dotenv import load_dotenv

from langchain.chains import LLMChain
from langchain.chat_models import ChatOpenAI
from langchain.agents import Tool, initialize_agent, AgentType
from langchain.prompts import PromptTemplate
from langchain.llms import OpenAI
from langchain.memory import ConversationBufferMemory

load_dotenv()

def fib(n):
    if n <= 1:
        return n
    else:
        return (fib(n-1) + fib(n-2))

def discord(message):
    prompt = PromptTemplate(
        template="""
            You are a message interpreter for Discord messages. Extract the recipient and core message content from: {message}.
            Output ONLY a valid JSON object with no extra text, following this structure:
            {{
              "recipient": "name_of_recipient",
              "message": "core_message_text"
            }}
        """, 
        input_variables=["message"],
    )

    llm = OpenAI(temperature=0)
    chain = LLMChain(llm=llm, prompt=prompt, verbose=True)
    response = chain.run(message)
    
    # 增加异常处理，避免JSON解析失败中断流程
    try:
        jsonResp = json.loads(response)
        return f"[Discord Tool Success] Sent '{jsonResp['message']}' to {jsonResp['recipient']}"
    except json.JSONDecodeError:
        return "[Discord Tool Error] Failed to parse message content"

tools = [
    Tool(
        name="Fibonacci",
        func=lambda n: str(fib(int(n))),
        description="Use when you want to calculate the nth fibonacci number"
    ),
    Tool(
        name="Discord",
        func=discord,
        description="Use when you need to send a message via Discord. Input should include recipient and message content in natural language."
    ),
]

memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
llm=ChatOpenAI(temperature=0, verbose=True)

agent_chain = initialize_agent(
    tools,
    llm,
    agent=AgentType.CHAT_CONVERSATIONAL_REACT_DESCRIPTION,
    memory=memory,
    verbose=True
)

response = agent_chain.run(input="Ask John on discord to grab some apple before coming home ")
print(f"response : {response}")

效果验证 #

修改后，Agent能正确识别工具的成功响应，最终会返回类似"Successfully sent the message to John via Discord."的正常结果，不再出现错误提示。

内容的提问来源于stack exchange，提问作者Dylan Grum's