如何在LangChain的RetrievalQA中添加记忆与多输入自定义提示词

要解决RetrievalQA无法同时支持多输入自定义提示词和对话记忆的问题，需要调整链的结构配置，确保额外参数（客户信息）和对话历史能正确传递到提示词中。以下是具体解决方案和修正后的代码：

关键问题分析 #

原代码存在两个核心问题：

• RetrievalQA的chain_type_kwargs中直接传入memory的方式不正确，RetrievalQA本身不支持这种配置方式，对话记忆需要通过正确的链绑定逻辑传递。
• 自定义提示词中的多输入参数（Customer_Name、Customer_State、Customer_Gender）需要被显式传递到QA链的调用上下文里，确保提示词能正确渲染。

修正后的实现代码 #

import openai
import numpy as np
import pandas as pd
import os
import json
from langchain.embeddings.sentence_transformer import SentenceTransformerEmbeddings
from langchain.vectorstores import Chroma
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.llms import OpenAI
from langchain.chains import RetrievalQA
from langchain.chains.question_answering import load_qa_chain
from langchain.document_loaders import UnstructuredFileLoader
from langchain.prompts import PromptTemplate
from langchain.memory import ConversationBufferMemory
from langchain.docstore.document import Document

# 加载PDF文档
loader = UnstructuredFileLoader("../Test.pdf", mode="elements")
documents = loader.load()

# 加载客户信息
with open('Customer_profile.json', 'r') as openfile:
    json_object = json.load(openfile)
cName = json_object['Customer_Name']
cState = json_object['Customer_State']
cGen = json_object['Customer_Gender']

# 定义带多输入的自定义提示词
prompt_template = """You are a Chat customer support agent.
Address the customer as Dear Mr. or Miss. depending on customer's gender followed by Customer's First Name.
Use the following customer related information (delimited by <cp></cp>) context (delimited by <ctx></ctx>) and the chat history (delimited by <hs></hs>) to answer the question at the end:
If you don't know the answer, just say that you don't know, don't try to make up an answer.
Below are the details of the customer:
<cp>
Customer's Name: {Customer_Name}
Customer's Resident State: {Customer_State}
Customer's Gender: {Customer_Gender}
</cp>
<ctx>
{context}
</ctx>
<hs>
{history}
</hs>
Question: {query}
Answer: """

PROMPT = PromptTemplate(
    template=prompt_template,
    input_variables=["history", "context", "query", "Customer_Name", "Customer_State", "Customer_Gender"]
)

# 文本分块与向量存储构建
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
texts = text_splitter.split_documents(documents)
embeddings = SentenceTransformerEmbeddings(model_name="all-MiniLM-L6-v2")
vectorDB = Chroma.from_documents(texts, embeddings)

# 配置对话记忆：只存储对话历史，输入key匹配query
memory = ConversationBufferMemory(
    memory_key="history",
    input_key="query",
    return_messages=False  # 直接返回字符串格式的历史，适配提示词
)

# 加载自定义QA链，使用stuff模式
qa_chain = load_qa_chain(
    llm=OpenAI(),
    chain_type="stuff",
    prompt=PROMPT,
    verbose=True
)

# 构建RetrievalQA，将自定义链传入，并绑定记忆
qa = RetrievalQA(
    combine_documents_chain=qa_chain,
    retriever=vectorDB.as_retriever(),
    verbose=True,
    memory=memory,
    return_source_documents=False
)

# 调用链，传入所有必要参数
response = qa({
    "query": "who's the client's friend?",
    "Customer_Gender": cGen,
    "Customer_State": cState,
    "Customer_Name": cName
})
print(response['result'])

# 后续对话示例，记忆会自动携带历史
response2 = qa({
    "query": "What's his contact information?",
    "Customer_Gender": cGen,
    "Customer_State": cState,
    "Customer_Name": cName
})
print(response2['result'])

核心修改说明 #

• 对话记忆的正确绑定：将memory直接作为RetrievalQA的初始化参数传入，而非放在chain_type_kwargs中，确保RetrievalQA能正确管理对话历史的存储与传递；设置return_messages=False，让记忆返回纯文本格式的对话历史，适配自定义提示词的渲染需求。
• 多输入参数的传递：调用链时必须显式传入所有自定义提示词中定义的输入变量，这些参数会被自动传递到提示词中完成渲染。
• 自定义QA链的加载：使用load_qa_chain加载自定义提示词生成符合需求的QA链，再将其传入RetrievalQA的combine_documents_chain参数，替代默认链类型配置，实现完全自定义的提示词逻辑。

内容的提问来源于stack exchange，提问作者Jason