Langchain中ConversationalRetrievalChain记忆功能失效问题求助

核心问题分析 #

1. 重复初始化记忆与链：Streamlit每次交互都会重新执行main函数，当前代码每次都会创建新的ConversationBufferMemory和ConversationalRetrievalChain，导致之前的会话记忆被清空。
2. 手动维护聊天记录与内置记忆冲突：ConversationalRetrievalChain已通过memory参数绑定记忆组件，无需手动传入chat_history参数；同时手动维护的st.session_state.chat_history与内置记忆逻辑重复，造成冲突。
3. 聊天记录存储格式错误：更新st.session_state.chat_history时，错误地将整个聊天记录列表嵌套追加，导致格式变为多层嵌套，无法被正确识别。

修复后的完整代码 #

import streamlit as st
import openai
import os
import pinecone
from dotenv import load_dotenv
from langchain.chat_models import AzureChatOpenAI
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import Pinecone
from streamlit_chat import message
from langchain.memory import ConversationBufferMemory
from langchain.chains import ConversationalRetrievalChain

# 加载环境变量
load_dotenv()

OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
OPENAI_DEPLOYMENT_ENDPOINT = os.getenv("OPENAI_DEPLOYMENT_ENDPOINT")
OPENAI_DEPLOYMENT_NAME = os.getenv("OPENAI_DEPLOYMENT_NAME")
OPENAI_MODEL_NAME = os.getenv("OPENAI_MODEL_NAME")
OPENAI_EMBEDDING_DEPLOYMENT_NAME = os.getenv("OPENAI_EMBEDDING_DEPLOYMENT_NAME")
OPENAI_EMBEDDING_MODEL_NAME = os.getenv("OPENAI_EMBEDDING_MODEL_NAME")
OPENAI_API_VERSION = os.getenv("OPENAI_API_VERSION")
OPENAI_API_TYPE = os.getenv("OPENAI_API_TYPE")

# Pinecone配置
PINECONE_API_KEY = os.getenv("PINECONE_API_KEY")
PINECONE_ENV = os.getenv("PINECONE_ENV")

# 初始化Azure OpenAI
openai.api_type = OPENAI_API_TYPE
openai.api_version = OPENAI_API_VERSION
openai.api_base = OPENAI_DEPLOYMENT_ENDPOINT
openai.api_key = OPENAI_API_KEY

st.set_page_config(
    page_title="Streamlit Chat - Demo",
    page_icon=":robot:"
)

def get_text():
    input_text = st.text_input("You: ","Who is John Doe?", key="input")
    return input_text 

def main():
    st.title('Scenario 2: Question Aswering on documents with langchain, pinecone and openai')
    st.markdown(
        """
        This scenario shows how to chat wih a txt file which was indexed in pinecone.
        """
    )

    pinecone.init(
        api_key=PINECONE_API_KEY,
        environment=PINECONE_ENV
    )
        
    # 初始化会话状态
    if 'generated' not in st.session_state:
        st.session_state['generated'] = []

    if 'past' not in st.session_state:
        st.session_state['past'] = []

    # 仅初始化一次记忆和链
    if 'chain' not in st.session_state:
        index_name = "default"
        embed = OpenAIEmbeddings(deployment=OPENAI_EMBEDDING_DEPLOYMENT_NAME, model=OPENAI_EMBEDDING_MODEL_NAME, chunk_size=1)
        retriever = Pinecone.from_existing_index(index_name, embed).as_retriever()
        
        llm = AzureChatOpenAI(
            openai_api_base=OPENAI_DEPLOYMENT_ENDPOINT,
            openai_api_version=OPENAI_API_VERSION,
            deployment_name=OPENAI_DEPLOYMENT_NAME,
            openai_api_key=OPENAI_API_KEY,
            openai_api_type=OPENAI_API_TYPE,
            model_name=OPENAI_MODEL_NAME,
            temperature=0)
        
        memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
        st.session_state['chain'] = ConversationalRetrievalChain.from_llm(llm, retriever, memory=memory)

    user_input = get_text()

    if user_input:
        # 直接使用已初始化的链，无需手动传入chat_history
        result = st.session_state['chain']({"question": user_input})
        st.session_state.past.append(user_input)
        st.session_state.generated.append(result['answer'])
      
    # 渲染聊天记录
    if st.session_state['generated']:
        for i in range(len(st.session_state['generated'])-1, -1, -1):
            message(st.session_state["generated"][i], key=str(i))
            message(st.session_state['past'][i], is_user=True, key=str(i) + '_user')

if __name__ == "__main__":
    main()

关键修复点说明 #

• 将memory和chain的初始化逻辑放到if 'chain' not in st.session_state:分支中，确保仅在首次加载时创建，避免每次交互重置记忆。
• 移除手动维护的chat_history相关冗余代码，直接依赖ConversationBufferMemory自动管理会话记忆。
• 调用链时仅传入question参数，无需额外传入chat_history，链会自动从绑定的记忆中获取上下文。

内容的提问来源于stack exchange，提问作者Luis Valencia