基于Python的自然语言处理系列（64）：Step-Back Prompting 问答技术-优快云博客

本文链接：https://blog.youkuaiyun.com/ljd939952281/article/details/146504214

介绍

在复杂问题的问答任务中，一种名为 Step-Back Prompting（回退提示）的方法能够提高回答的准确性。该技术的核心思想是先提出一个更通用的“回退”问题，再利用它来辅助回答原始问题。这一方法可以与检索增强问答（Retrieval-Augmented Question Answering，RAG）结合，以提升信息检索的质量。

相关论文可查看：Step-Back Prompting 论文

一篇优秀的相关博客文章：Cobus Greyling - Step-Back Prompting

在本教程中，我们将复现 Step-Back Prompting 的实现，并调整提示词使其更适用于聊天模型。

1. 环境准备

from langchain.chat_models import ChatOpenAI
from langchain.prompts import ChatPromptTemplate, FewShotChatMessagePromptTemplate
from langchain.schema.output_parser import StrOutputParser
from langchain.schema.runnable import RunnableLambda

2. 创建 Few-Shot 示例

为了帮助模型更好地理解 Step-Back Prompting，我们先提供一些示例：

examples = [
    {"input": "Could the members of The Police perform lawful arrests?",
     "output": "what can the members of The Police do?"},
    {"input": "Jan Sindel’s was born in what country?",
     "output": "what is Jan Sindel’s personal history?"},
]

example_prompt = ChatPromptTemplate.from_messages([
    ("human", "{input}"),
    ("ai", "{output}"),
])

few_shot_prompt = FewShotChatMessagePromptTemplate(
    example_prompt=example_prompt,
    examples=examples,
)

3. 构造 Step-Back Prompting 生成器

我们创建一个 ChatPromptTemplate，将 Few-Shot 示例和系统指令结合起来：

prompt = ChatPromptTemplate.from_messages([
    ("system", """
    You are an expert at world knowledge. Your task is to step back and paraphrase a question to a more generic step-back question, which is easier to answer.
    Here are a few examples:
    """),
    few_shot_prompt,
    ("user", "{question}"),
])

question_gen = prompt | ChatOpenAI(temperature=0) | StrOutputParser()

4. 测试 Step-Back Prompting

question = "was chatgpt around while trump was president?"
print(question_gen.invoke({"question": question}))

5. 检索增强问答 (RAG) 结合 Step-Back Prompting

我们使用 DuckDuckGo 搜索 API 来检索相关信息：

from langchain.utilities import DuckDuckGoSearchAPIWrapper

search = DuckDuckGoSearchAPIWrapper(max_results=4)

def retriever(query):
    return search.run(query)

# 使用普通问题进行检索
retriever(question)

# 使用 Step-Back 生成的问题进行检索
retriever(question_gen.invoke({"question": question}))

6. 组合 Step-Back 和问答系统

我们从 LangChain Hub 加载标准的问答模板，并结合 Step-Back 机制进行检索和回答。

from langchain import hub

response_prompt = hub.pull("langchain-ai/stepback-answer")

chain = (
    {
        "normal_context": RunnableLambda(lambda x: x["question"]) | retriever,
        "step_back_context": question_gen | retriever,
        "question": lambda x: x["question"],
    }
    | response_prompt
    | ChatOpenAI(temperature=0)
    | StrOutputParser()
)

运行 Step-Back 问答系统

chain.invoke({"question": question})

7. 基线实验（不使用 Step-Back Prompting）

我们也可以进行基线测试，仅使用普通检索进行问答。

response_prompt_template = """
You are an expert of world knowledge. I am going to ask you a question. Your response should be comprehensive and not contradicted with the following context if they are relevant. Otherwise, ignore them if they are not relevant.

{normal_context}

Original Question: {question}
Answer:"""

response_prompt = ChatPromptTemplate.from_template(response_prompt_template)

chain_baseline = (
    {
        "normal_context": RunnableLambda(lambda x: x["question"]) | retriever,
        "question": lambda x: x["question"],
    }
    | response_prompt
    | ChatOpenAI(temperature=0)
    | StrOutputParser()
)

# 运行基线模型
chain_baseline.invoke({"question": question})