從0開始學(xué)RAG之Query分解
原文地址：https://zhuanlan.zhihu.com/p/685746861
類型: 技術(shù)分享
本文為 @lucas大叔 投稿原創(chuàng)轉(zhuǎn)載！如有侵權(quán)，麻煩告知?jiǎng)h除！
Query分解是通過將問題分解為子問題來改善問答效果的策略，分兩條實(shí)現(xiàn)路徑：（1）序列求解，將上一個(gè)子問題的答案與當(dāng)前問題一并扔給LLM生成答案，再把當(dāng)前生成答案與下一個(gè)子問題一起給LLM生成答案，直到最后一個(gè)子問題生成最終答案；（2）并行的獨(dú)立回答問題，然后將多路答案合并為最終答案。
下面的示意圖很好地詮釋了兩種技術(shù)實(shí)現(xiàn)路徑的原理。
序列求解的思想來自least-to-most prompting和IRCoT這兩篇論文。
在least-to-most prompting中，作者認(rèn)為CoT提示在自然語(yǔ)言推理任務(wù)上表現(xiàn)搶眼，但在解決比給的樣例更難的問題時(shí)就乏善可陳了。為了解決easy-to-hard泛化問題，大佬們提出了least-to-most prompting。分兩步解決問題：首先將復(fù)雜問題分解成一系列更容易的子問題，然后依次求解這些子問題，用先前解決的子問題答案助力當(dāng)前子問題的解決。這兩步都是通過few-shot prompting來實(shí)現(xiàn)，不需要訓(xùn)練或者微調(diào)。
在IRCoT中作者提出了類似的想法，對(duì)多步問答在CoT中交錯(cuò)檢索，使用檢索到的結(jié)果改善CoT。
更多技術(shù)細(xì)節(jié)，大家可以參考上述的兩篇論文。下面我們轉(zhuǎn)入代碼實(shí)戰(zhàn)環(huán)節(jié)。
首先寫提示模板，將問題分解為若干個(gè)子問題。
from langchain.prompts import ChatPromptTemplate# Decompositiontemplate = """You are a helpful assistant that generates multiple sub-questions related to an input question. \nThe goal is to break down the input into a set of sub-problems / sub-questions that can be answers in isolation. \nGenerate multiple search queries related to: {question} \nOutput (3 queries):"""prompt_decomposition = ChatPromptTemplate.from_template(template)
構(gòu)造分解問題的chain，并將問題分解為幾個(gè)子問題。
from langchain_openai import ChatOpenAIfrom langchain_core.output_parsers import StrOutputParser# LLMllm = ChatOpenAI(temperature=0)# Chaingenerate_queries_decomposition = ( prompt_decomposition | llm | StrOutputParser() | (lambda x: x.split("\n")))# Runquestion = "What are the main components of an LLM-powered autonomous agent system?"questions = generate_queries_decomposition.invoke({"question":question})
Answer recursively
如上面序列求解流程所示，構(gòu)建序列回答的prompt模板，基于context和qa對(duì)回答問題。
# Prompttemplate = """Here is the question you need to answer:\n --- \n {question} \n --- \nHere is any available background question + answer pairs:\n --- \n {q_a_pairs} \n --- \nHere is additional context relevant to the question:\n --- \n {context} \n --- \nUse the above context and any background question + answer pairs to answer the question: \n {question}"""decomposition_prompt = ChatPromptTemplate.from_template(template)初始化q_a_pairs為空，查詢第一個(gè)問題時(shí)，沒有qa對(duì)信息。從第二個(gè)問題開
始，除了當(dāng)前的問題，還有前面所有輪的qa對(duì)信息，再加上當(dāng)前問題檢索回來的context一起給LLM得到答案。
from operator import itemgetterfrom langchain_core.output_parsers import StrOutputParserdef format_qa_pair(question, answer):"""Format Q and A pair"""formatted_string = ""formatted_string += f"Question: {question}\nAnswer: {answer}\n\n"return formatted_string.strip()# llmllm = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0)#q_a_pairs = ""for q in questions:rag_chain = ({"context": itemgetter("question") | retriever,"question": itemgetter("question"),"q_a_pairs": itemgetter("q_a_pairs")}| decomposition_prompt| llm| StrOutputParser())answer = rag_chain.invoke({"question":q,"q_a_pairs":q_a_pairs})q_a_pair = format_qa_pair(q,answer)q_a_pairs = q_a_pairs + "\n---\n"+ q_a_pair
Answer individually
相比于序列回答，并行獨(dú)立回答邏輯簡(jiǎn)單的多。每個(gè)子問題各自調(diào)用LLM回答自己的提問，匯總得到的答案。
# Answer each sub-question individuallyfrom langchain import hubfrom langchain_core.prompts import ChatPromptTemplatefrom langchain_core.runnables import RunnablePassthrough, RunnableLambdafrom langchain_core.output_parsers import StrOutputParserfrom langchain_openai import ChatOpenAI# RAG promptprompt_rag = hub.pull("rlm/rag-prompt")def retrieve_and_rag(question,prompt_rag,sub_question_generator_chain):"""RAG on each sub-question"""# Use our decomposition /sub_questions = sub_question_generator_chain.invoke({"question":question})# Initialize a list to hold RAG chain resultsrag_results = []for sub_question in sub_questions:# Retrieve documents for each sub-questionretrieved_docs = retriever.get_relevant_documents(sub_question)# Use retrieved documents and sub-question in RAG chainanswer = (prompt_rag | llm | StrOutputParser()).invoke({"context": retrieved_docs,"question": sub_question})rag_results.append(answer)return rag_results,sub_questions# Wrap the retrieval and RAG process in a RunnableLambda for integration into a chainanswers, questions = retrieve_and_rag(question, prompt_rag, generate_queries_decomposition)
提示模板的指令也簡(jiǎn)單粗暴，告訴模型這是一組QA對(duì)，你用它們來合成原始問題的答案吧！
def format_qa_pairs(questions, answers):"""Format Q and A pairs"""formatted_string = ""for i, (question, answer) in enumerate(zip(questions, answers), start=1):formatted_string += f"Question {i}: {question}\nAnswer {i}: {answer}\n\n"return formatted_string.strip()context = format_qa_pairs(questions, answers)# Prompttemplate = """Here is a set of Q+A pairs:{context}Use these to synthesize an answer to the question: {question}"""prompt = ChatPromptTemplate.from_template(template)final_rag_chain = (prompt| llm| StrOutputParser())final_rag_chain.invoke({"context":context,"question":question})