84 LlamaIndex 相关性评估器使用指南 （响应和查询的相关性）

LlamaIndex 相关性评估器使用指南

在大型语言模型（LLM）的应用中，确保生成的响应与查询相关且准确是至关重要的。LlamaIndex 提供了 RelevancyEvaluator 模块，帮助开发者评估响应和源节点是否与查询匹配。本文将详细介绍如何使用 RelevancyEvaluator 进行相关性评估。

安装依赖

首先，我们需要安装必要的依赖包：

%pip install llama-index-llms-openai pandas[jinja2] spacy

设置日志

为了更好地调试和查看运行过程中的信息，我们设置日志：

import logging
import sys

logging.basicConfig(stream=sys.stdout, level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler(stream=sys.stdout))

导入必要的模块

接下来，导入 LlamaIndex 和其他必要的模块：

from llama_index.core import (
    TreeIndex,
    VectorStoreIndex,
    SimpleDirectoryReader,
    Response,
)
from llama_index.llms.openai import OpenAI
from llama_index.core.evaluation import RelevancyEvaluator
from llama_index.core.node_parser import SentenceSplitter
import pandas as pd

pd.set_option("display.max_colwidth", 0)

使用 GPT-3 和 GPT-4 进行评估

我们使用 GPT-3 和 GPT-4 作为评估模型：

# gpt-3 (davinci)
gpt3 = OpenAI(temperature=0, model="gpt-3.5-turbo")

# gpt-4
gpt4 = OpenAI(temperature=0, model="gpt-4")

evaluator = RelevancyEvaluator(llm=gpt3)
evaluator_gpt4 = RelevancyEvaluator(llm=gpt4)

加载文档并创建向量索引

加载测试数据并创建向量索引：

documents = SimpleDirectoryReader("./test_wiki_data").load_data()
splitter = SentenceSplitter(chunk_size=512)
vector_index = VectorStoreIndex.from_documents(
    documents, transformations=[splitter]
)

定义显示评估结果的函数

定义一个函数，用于在 Jupyter Notebook 中显示评估结果：

from llama_index.core.evaluation import EvaluationResult

def display_eval_df(
    query: str, response: Response, eval_result: EvaluationResult
) -> None:
    eval_df = pd.DataFrame(
        {
            "Query": query,
            "Response": str(response),
            "Source": response.source_nodes[0].node.text[:1000] + "...",
            "Evaluation Result": "Pass" if eval_result.passing else "Fail",
            "Reasoning": eval_result.feedback,
        },
        index=[0],
    )
    eval_df = eval_df.style.set_properties(
        **{
            "inline-size": "600px",
            "overflow-wrap": "break-word",
        },
        subset=["Response", "Source"]
    )
    display(eval_df)

评估响应

运行查询并评估响应：

query_str = "What battles took place in New York City in the American Revolution?"
query_engine = vector_index.as_query_engine()
response_vector = query_engine.query(query_str)
eval_result = evaluator_gpt4.evaluate_response(
    query=query_str, response=response_vector
)
display_eval_df(query_str, response_vector, eval_result)

query_str = "What are the airports in New York City?"
query_engine = vector_index.as_query_engine()
response_vector = query_engine.query(query_str)
eval_result = evaluator_gpt4.evaluate_response(
    query=query_str, response=response_vector
)
display_eval_df(query_str, response_vector, eval_result)

query_str = "Who is the mayor of New York City?"
query_engine = vector_index.as_query_engine()
response_vector = query_engine.query(query_str)
eval_result = evaluator_gpt4.evaluate_response(
    query=query_str, response=response_vector
)
display_eval_df(query_str, response_vector, eval_result)

评估源节点

评估返回的源节点，确定哪些源节点实际包含给定查询的答案：

from typing import List

def display_eval_sources(
    query: str, response: Response, eval_result: List[str]
) -> None:
    sources = [s.node.get_text() for s in response.source_nodes]
    eval_df = pd.DataFrame(
        {
            "Source": sources,
            "Eval Result": eval_result,
        },
    )
    eval_df.style.set_caption(query)
    eval_df = eval_df.style.set_properties(
        **{
            "inline-size": "600px",
            "overflow-wrap": "break-word",
        },
        subset=["Source"]
    )

    display(eval_df)

query_str = "What are the airports in New York City?"
query_engine = vector_index.as_query_engine(
    similarity_top_k=3, response_mode="no_text"
)
response_vector = query_engine.query(query_str)
eval_source_result_full = [
    evaluator_gpt4.evaluate(
        query=query_str,
        response=response_vector.response,
        contexts=[source_node.get_content()],
    )
    for source_node in response_vector.source_nodes
]
eval_source_result = [
    "Pass" if result.passing else "Fail" for result in eval_source_result_full
]
display_eval_sources(query_str, response_vector, eval_source_result)

query_str = "Who is the mayor of New York City?"
query_engine = vector_index.as_query_engine(
    similarity_top_k=3, response_mode="no_text"
)
eval_source_result_full = [
    evaluator_gpt4.evaluate(
        query=query_str,
        response=response_vector.response,
        contexts=[source_node.get_content()],
    )
    for source_node in response_vector.source_nodes
]
eval_source_result = [
    "Pass" if result.passing else "Fail" for result in eval_source_result_full
]
display_eval_sources(query_str, response_vector, eval_source_result)

总结

通过 LlamaIndex 的 RelevancyEvaluator 模块，我们可以有效地评估查询引擎的响应和源节点是否与查询相关。这不仅有助于确保生成的内容准确无误，还能提升用户体验和系统的整体性能。希望本文能帮助你更好地理解和应用 LlamaIndex 的相关性评估功能。如果你有任何问题或需要进一步的帮助，请随时联系我们。祝你在 LLM 开发的道路上取得成功！