让模型评估模型：构建双代理RAG评估系统的步骤解析

在当前大语言模型(LLM)应用开发的背景下,一个关键问题是如何评估模型输出的准确性。我们需要确定哪些评估指标能够有效衡量提示(prompt)的效果,以及在多大程度上需要对提示进行优化。

为解决这一问题,我们将介绍一个基于双代理的RAG(检索增强生成)评估系统。该系统使用生成代理和反馈代理,基于预定义的测试集对输出进行评估。或者更简单的说，我们使用一个模型来评估另外一个模型的输出。

在本文中将详细介绍如何构建这样一个RAG评估系统,并展示基于四种提示工程技术的不同结果,包括ReAct、思维链(Chain of Thought)、自一致性(Self-Consistency)和角色提示(Role Prompting)。

以下是该项目的整体架构图:

数据收集与摄入

此部分在 ingestion.py 中实现

数据收集过程使用了三篇文章作为源数据。在加载和分割数据后,我们对文本进行嵌入,并将嵌入向量存储在FAISS中。FAISS(Facebook AI Similarity Search)是由Meta开发的开源库,用于高效进行密集向量的相似性搜索和聚类。

以下是实现代码:

 urls= [  
     "https://medium.com/@fareedkhandev/prompt-engineering-complete-guide-2968776f0431",  
     "https://medium.com/@researchgraph/prompt-engineering-21112dbfc789",  
     "https://blog.fabrichq.ai/what-is-prompt-engineering-a-detailed-guide-with-examples-4d3cbbd53792"  
 ]  
 loader=WebBaseLoader(urls)  
 # 文本分割器  
 text_splitter=RecursiveCharacterTextSplitter(  
     chunk_size=1000, chunk_overlap=20  
 )  
 documents=loader.load_and_split(text_splitter)  
 
 # LLM  
 embedder_llm=OpenAIModel().embed_model()  
 
 # 对文档块进行嵌入  
 vectorstore=FAISS.from_documents(documents, embedder_llm)  
 vectorstore.save_local("faiss_embed")  
 print("===== 数据摄入完成 ===== ")

创建测试集

此部分在 create_test_set.py 中实现

测试集的构建使用了Giskard工具。Giskard是一个开源工具,专为测试和改进机器学习模型而设计。它使用户能够创建、运行和自动化测试,以评估模型的性能、公平性和稳健性。

实现代码如下:

 fromlangchain_community.document_loadersimportWebBaseLoader  
 fromlangchain.text_splitterimportRecursiveCharacterTextSplitter  
 # 用于构建测试集  
 fromgiskard.ragimportKnowledgeBase, generate_testset  
 # 数据框  
 importpandasaspd  
 fromLLM.modelsimportOpenAIModel  
 
 if__name__=='__main__':  
     urls= [  
         "https://medium.com/@fareedkhandev/prompt-engineering-complete-guide-2968776f0431",  
         "https://medium.com/@researchgraph/prompt-engineering-21112dbfc789",  
         "https://blog.fabrichq.ai/what-is-prompt-engineering-a-detailed-guide-with-examples-4d3cbbd53792"  
     ]  
     loader=WebBaseLoader(urls)  
     # 文本分割器  
     text_splitter=RecursiveCharacterTextSplitter(  
         chunk_size=1000, chunk_overlap=20  
     )  
     documents=loader.load_and_split(text_splitter)  
 
     df=pd.DataFrame([doc.page_contentfordocindocuments], columns=["text"])  
     print(df.head(10))  
 
     ## 将数据框添加到giskard KnowledgeBase  
     knowledge_base=KnowledgeBase(df)  
 
     # 生成测试集  
     test_set=generate_testset(  
         knowledge_base,  
         num_questions=10,  
         agent_description="A chatbot answering question about prompt engineering"  
     )  
     test_set.save("test-set.jsonl")

由于文本太多，生成的样例就不显示了

答案检索

此部分在 generation.py 中实现

本文的第一个流程是生成流程。我们从FAISS检索数据。实现代码如下：

 generate_llm=OpenAIModel().generate_model()  
 embedder_llm=OpenAIModel().embed_model()  
 vectorstore=FAISS.load_local("faiss_embed", embedder_llm, allow_dangerous_deserialization=True)  
 
 retrieval_qa_chat_prompt= (retrieval)  
 
 prompt=ChatPromptTemplate.from_messages(  
     [  
         ("system", retrieval_qa_chat_prompt),  
         ("human", "{input}"),  
     ]  
 )
 combine_docs_chain=create_stuff_documents_chain(generate_llm, prompt)  
 retrival_chain=create_retrieval_chain(  
     retriever=vectorstore.as_retriever(),  
     combine_docs_chain=combine_docs_chain  
 )

评估

此部分在 evaluation.py 中实现

评估过程中向LLM提供三个输入：问题、AI答案（第一个LLM的输出）和实际答案（从测试集中检索）。实现代码如下：

 defRAG_eval(question, AI_answer, Actual_answer, prompt):  
     evaluation_prompt_template=PromptTemplate(  
         input_variables=[  
             "question", "AI_answer", "Actual_answer"  
         ],  
         template=prompt  
     )  
     generate_llm=OpenAIModel().generate_model()  
 
     optimization_chain=evaluation_prompt_template|generate_llm|StrOutputParser()  
 
     result_optimization=optimization_chain.invoke(  
         {"question": question, "AI_answer": AI_answer, "Actual_answer": Actual_answer})  
     returnresult_optimization

链接整合

此部分在 main.py 中实现

主文件遍历测试数据，使用问题作为第一个LLM的输入。然后将第一个LLM的输出用作第二个LLM的输入。实现代码如下：

 foritemindata:  
     question= {"input": item['question']}  
     # 生成回答  
     result=retrival_chain.invoke(input=question)  
     AI_answer=result['answer']  
     # 获取实际答案  
     Actual_answer=item['reference_answer']  
 
     # 将所有内容提供给第二个LLM  
     Evaluation=RAG_eval(  
         question=question,  
         AI_answer=AI_answer,  
         Actual_answer=Actual_answer,  
         prompt=evaluation_self_consistency_prompting  
     )  
 
     print(f"AI_answer:{AI_answer}")  
     print(Evaluation)

实验结果

评估组件采用了四种不同的提示工程技术：

思维链（Chain of Thought）
ReAct
角色提示（Role Prompting）
自一致性（Self-Consistency）

以下是基于这四种不同提示技术的评估代理对测试集第一个问题的输出示例：

问题： What is the purpose of incorporating knowledge in prompt engineering?

实际答案： Incorporating knowledge or information in prompt engineering enhances the model’s prediction accuracy. By providing relevant knowledge or information related to the task at hand, the model can leverage this additional context to make more accurate predictions. This technique enables the model to tap into external resources or pre-existing knowledge to improve its understanding and generate more informed responses

**AI答案：**Incorporating knowledge in prompt engineering enhances the quality of responses by guiding AI models to provide not just answers, but also relevant context and insights. This leads to more informative and meaningful interactions, improving user experience and understanding.

使用思维链输出的**评估结果：**The student’s answer correctly identifies that incorporating knowledge enhances the quality of responses and improves user experience. However, it lacks emphasis on the model’s prediction accuracy and the importance of leveraging external resources or pre-existing knowledge, which are key components of the actual answer. This omission affects the completeness and correctness of the response.

使用ReAct输出的评估结果**：**The student’s answer correctly identifies that incorporating knowledge enhances the quality of responses and improves user experience. However, it lacks emphasis on the specific aspect of prediction accuracy and the importance of leveraging external resources or pre-existing knowledge, which are key components of the actual answer. This omission affects the completeness of the response, leading to a lower evaluation percentage.

使用角色提示输出的评估结果**：**The student’s response accurately captures the essence of incorporating knowledge in prompt engineering by emphasizing the enhancement of response quality and user experience. However, it lacks specific mention of prediction accuracy and the model’s ability to leverage external resources, which are key aspects of the actual response.

使用自一致性输出的评估结果**：**The student’s answer captures the essence of enhancing the quality of responses through knowledge incorporation, but it lacks specific mention of prediction accuracy and the model’s ability to leverage external resources. The initial evaluation was slightly optimistic, but upon reevaluation, it became clear that the answer did not fully align with the actual answer’s emphasis on prediction accuracy and context utilization