修复LangChain调用MistralAI Agent多轮对话的400错误

错误信息 #

HTTPStatusError: Error response 400 while fetching https://api.mistral.ai/v1/chat/completions: {"object":"error","message":"Expected last role User or Tool (or Assistant with prefix True) for serving but got assistant","type":"invalid_request_message_order","param":null,"code":"3230"}
During task with name 'model' and id '61e16c9d-830c-9496-12da-acafb10a9a50'

用户代码 #

from langchain.chat_models import init_chat_model

model = init_chat_model(
    "mistralai:mistral-small-latest",
    temperature=0.7,
    timeout=10,
    max_tokens=1000
)

from langchain.agents import create_agent

agent = create_agent(
    model = model,
    system_prompt = system_prompt,
    tools = [get_user_location, get_weather_for_location],
    context_schema = Context,
    response_format = ResponseFormat,
    checkpointer = checkpointer
)

def invoke_agent_stateful(agent, user_input: str, user_id: str):
    config = {"configurable": {"thread_id": f"{user_id}"}}
    response = agent.invoke(
        {"messages": [{"role": "user", "content": f"{user_input}"}]},
        config=config,
        context=Context(user_id=user_id)
    )
    return response['structured_response']

invoke_agent_stateful(agent, user_input="what is the weather?", user_id="1") # 首次调用输出正常

invoke_agent_stateful(agent, user_input="where am i?", user_id="1") # 调用出现错误

问题根源 #

Mistral API要求对话历史的角色顺序必须严格交替（User ↔ Assistant），或在工具调用场景下遵循User → Tool → Assistant的顺序。第二次调用时，对话历史的最后一条重复出现了Assistant角色，导致API返回格式错误。

修复方案 #

1. 保留完整Agent响应，确保状态正确更新 #

你的函数只返回了结构化响应，但Agent的完整输出包含更新对话历史的关键信息。修改函数，让checkpointer自动处理状态更新：

def invoke_agent_stateful(agent, user_input: str, user_id: str):
    config = {"configurable": {"thread_id": f"{user_id}"}}
    # 保留完整响应，确保checkpointer正确保存对话状态
    full_response = agent.invoke(
        {"messages": [{"role": "user", "content": f"{user_input}"}]},
        config=config,
        context=Context(user_id=user_id)
    )
    # 按需提取结构化响应
    return full_response['structured_response']

2. 显式包装Agent的消息历史管理 #

如果checkpointer配置存在问题，显式用RunnableWithMessageHistory包装Agent，确保对话历史顺序正确：

from langchain_core.runnables.history import RunnableWithMessageHistory
from langchain_core.chat_history import InMemoryChatMessageHistory

# 内存存储对话历史，实际场景可替换为数据库等持久化方案
chat_history_store = {}
def get_session_history(session_id: str):
    if session_id not in chat_history_store:
        chat_history_store[session_id] = InMemoryChatMessageHistory()
    return chat_history_store[session_id]

# 用消息历史包装Agent
agent_with_history = RunnableWithMessageHistory(
    agent,
    get_session_history,
    input_messages_key="messages",
    history_messages_key="history",
)

def invoke_agent_stateful(agent, user_input: str, user_id: str):
    config = {"configurable": {"session_id": f"{user_id}"}}
    response = agent_with_history.invoke(
        {"messages": [{"role": "user", "content": f"{user_input}"}]},
        config=config,
        context=Context(user_id=user_id)
    )
    return response['structured_response']

3. 验证对话历史格式 #

在调用前可以打印当前对话历史，确认角色顺序是否正确：

# 示例：获取并打印对话历史
history = get_session_history("1")
print([msg.dict() for msg in history.messages])

确保输出的角色顺序是user → assistant → user...，没有连续的assistant角色。

内容的提问来源于stack exchange，提问作者Gokul Krishna Balaji