293 lines
11 KiB
Python
293 lines
11 KiB
Python
"""
|
||
答案正确性评判工具
|
||
|
||
此模块用于评判问题的新旧回答是否正确，通过与标准答案（Wiki内容）进行比较，
|
||
或者在没有标准答案的情况下比较新旧回答的差异。
|
||
|
||
用法示例:
|
||
judge = AnswerCorrectnessJudge()
|
||
judge.process()
|
||
"""
|
||
|
||
import pandas as pd
|
||
from urllib.parse import unquote
|
||
from rag2_0.tool.WikijsTool import WikijsTool
|
||
from rag2_0.tool.html_to_md import convert_html_to_md
|
||
from rag2_0.tool.ModelTool import OpenAiLLM
|
||
from dotenv import load_dotenv
|
||
import os
|
||
from tqdm import tqdm
|
||
load_dotenv()
|
||
|
||
class AnswerCorrectnessJudge:
|
||
"""
|
||
答案正确性评判工具类
|
||
|
||
用于评估问题的新旧回答是否正确，可以通过与标准答案（Wiki内容）进行比较，
|
||
或者在没有标准答案的情况下比较新旧回答的差异。
|
||
"""
|
||
|
||
def __init__(self, wiki_excel_path="/data/Rag2_0/data/excel/部分提问_软件名称明确.xlsx",
|
||
answer_excel_path="/data/Rag2_0/data/excel/主网软件提问_对比结果.xlsx",
|
||
output_path="/data/Rag2_0/data/excel/主网软件提问回答_判断结果.xlsx"):
|
||
"""
|
||
初始化答案正确性评判工具
|
||
|
||
参数:
|
||
wiki_excel_path (str): Wiki Excel文件路径
|
||
answer_excel_path (str): 答案对比Excel文件路径
|
||
output_path (str): 输出Excel文件路径
|
||
"""
|
||
self.wiki_excel_path = wiki_excel_path
|
||
self.answer_excel_path = answer_excel_path
|
||
self.output_path = output_path
|
||
|
||
# 读取Excel文件
|
||
self.wiki_excel = pd.read_excel(self.wiki_excel_path)
|
||
self.answer_excel = pd.read_excel(self.answer_excel_path)
|
||
|
||
# 初始化LLM
|
||
self.api_key = os.getenv("OPENAI_API_KEY")
|
||
self.base_url = os.getenv("OPENAI_API_BASE")
|
||
self.model = os.getenv("LLM_MODEL_NAME")
|
||
|
||
if not all([self.api_key, self.base_url, self.model]):
|
||
raise ValueError("请设置 OPENAI_API_KEY, OPENAI_API_BASE, 和 LLM_MODEL_NAME 环境变量")
|
||
|
||
self.openai_llm = OpenAiLLM(api_key=self.api_key, base_url=self.base_url, model=self.model)
|
||
|
||
def find_wiki_link(self, query) -> str | None:
|
||
"""
|
||
根据查询（对应wiki_excel中的新提问列）找出对应的词条链接
|
||
|
||
参数:
|
||
query (str): 查询内容，对应wiki_excel中的新提问列
|
||
|
||
返回:
|
||
str: 对应的词条链接，如果没有找到则返回None
|
||
"""
|
||
# 确保query不为空
|
||
if not query or pd.isna(query):
|
||
return None
|
||
|
||
# 在"新提问"列中查找匹配的行
|
||
matched_rows = self.wiki_excel[self.wiki_excel['新提问'] == query]
|
||
|
||
# 如果找到了匹配的行，返回对应的词条链接
|
||
if not matched_rows.empty:
|
||
return matched_rows.iloc[0]['对应词条链接']
|
||
|
||
# 如果没有完全匹配，尝试部分匹配
|
||
# 去除软件名称部分（如果有）
|
||
query_parts = query.split(',', 1)
|
||
if len(query_parts) > 1:
|
||
clean_query = query_parts[1].strip()
|
||
|
||
# 在"提问"列中查找包含清理后查询的行
|
||
for idx, row in self.wiki_excel.iterrows():
|
||
if pd.notna(row['提问']) and clean_query in row['提问']:
|
||
return row['对应词条链接']
|
||
|
||
return None
|
||
|
||
def get_wiki_content(self, link) -> str:
|
||
"""
|
||
获取词条链接的内容
|
||
|
||
参数:
|
||
link (str): 词条链接
|
||
|
||
返回:
|
||
str: 链接内容，如果获取失败则返回错误信息
|
||
"""
|
||
try:
|
||
if not link or pd.isna(link):
|
||
return "链接为空或无效"
|
||
# 移除域名部分,只保留路径
|
||
path = link.split('/', 3)[-1]
|
||
decoded_path = unquote(path)
|
||
path_parts = decoded_path.split('/')
|
||
doc_path = "/".join(path_parts[1:])
|
||
wiki_doc = WikijsTool.get_all_doc_by_path(path=doc_path, path_is_dir=False)
|
||
html_content = WikijsTool.query_doc_info(wiki_doc[0]["id"]).get('content')
|
||
if not html_content:
|
||
return "获取内容失败"
|
||
|
||
options = {"heading_style": '', "keep_inline_images_in": ["figure", "img"], "escape_asterisks": True}
|
||
new_content = (html_content.replace("h6>", "h7>")
|
||
.replace("h5>", "h6>")
|
||
.replace("h4>", "h5>")
|
||
.replace("h3>", "h4>")
|
||
.replace("h2>", "h3>")
|
||
.replace("h1>", "h2>"))
|
||
# 将HTML内容转换为Markdown
|
||
markdown_content = convert_html_to_md(new_content, "", **options)
|
||
markdown_content = f"# {path_parts[-1]}\n\n{markdown_content}"
|
||
return markdown_content
|
||
|
||
except Exception as e:
|
||
raise RuntimeError(f"获取词条内容失败: {str(e)}") from e
|
||
|
||
def create_prompt(self, standard_answer: str, answer_to_check: str) -> str:
|
||
"""
|
||
创建用于评判答案的prompt
|
||
|
||
参数:
|
||
standard_answer (str): 标准答案
|
||
answer_to_check (str): 需要检查的答案
|
||
|
||
返回:
|
||
str: 格式化的prompt
|
||
"""
|
||
return f"""请作为一个专业的答案评判专家，评估以下回答与标准答案的匹配程度。
|
||
|
||
标准答案：
|
||
{standard_answer}
|
||
|
||
待评估的回答：
|
||
{answer_to_check}
|
||
|
||
请仔细分析两个答案的内容，并给出你的判断。只需要回答"正确"或"错误"，不需要其他解释。
|
||
如果待评估的回答与标准答案在核心内容和关键信息(步骤)上一致，即使表达方式不同，也应判定为"正确"。
|
||
如果待评估的回答存在明显的错误信息或重要信息缺失，应判定为"错误"。
|
||
|
||
请严格按以下格式输出：【正确】或【错误】："""
|
||
|
||
def judge_old_answer(self, standard_answer: str, old_answer: str) -> bool | None:
|
||
"""
|
||
调用LLM判断旧回答是否正确
|
||
|
||
参数:
|
||
standard_answer (str): 标准答案（来自Wiki）
|
||
old_answer (str): 旧流程的回答
|
||
|
||
返回:
|
||
bool | None: 判断结果，True表示正确，False表示错误，None表示判断失败
|
||
"""
|
||
prompt = self.create_prompt(standard_answer, old_answer)
|
||
try:
|
||
response = self.openai_llm.invoke(prompt)
|
||
return "正确" in response.content
|
||
except Exception as e:
|
||
return None
|
||
|
||
def judge_new_answer(self, standard_answer: str, new_answer: str) -> bool | None:
|
||
"""
|
||
调用LLM判断新回答是否正确
|
||
|
||
参数:
|
||
standard_answer (str): 标准答案（来自Wiki）
|
||
new_answer (str): 新流程的回答
|
||
|
||
返回:
|
||
bool | None: 判断结果，True表示正确，False表示错误，None表示判断失败
|
||
"""
|
||
prompt = self.create_prompt(standard_answer, new_answer)
|
||
try:
|
||
response = self.openai_llm.invoke(prompt)
|
||
return "正确" in response.content
|
||
except Exception as e:
|
||
return None
|
||
|
||
def judge_by_standard_answer(self, standard_answer: str, old_answer: str, new_answer: str) -> str | None:
|
||
"""
|
||
综合判断新旧回答的正确性
|
||
|
||
参数:
|
||
standard_answer (str): 标准答案（来自Wiki）
|
||
old_answer (str): 旧流程的回答
|
||
new_answer (str): 新流程的回答
|
||
|
||
返回:
|
||
str | None: 包含新旧回答判断结果的字符串，None表示判断失败
|
||
"""
|
||
old_result = self.judge_old_answer(standard_answer, old_answer)
|
||
new_result = self.judge_new_answer(standard_answer, new_answer)
|
||
if old_result is None or new_result is None:
|
||
return None
|
||
if new_result and old_result:
|
||
return "新旧答案均正确"
|
||
elif new_result and not old_result:
|
||
return "新答案正确"
|
||
elif not new_result and old_result:
|
||
return "旧答案正确"
|
||
else:
|
||
return "新旧答案均错误"
|
||
|
||
def judge_answer_diff(self, old_answer: str, new_answer: str) -> str | None:
|
||
"""
|
||
判断新旧回答是否存在较大差异
|
||
|
||
参数:
|
||
old_answer (str): 旧流程的回答
|
||
new_answer (str): 新流程的回答
|
||
|
||
返回:
|
||
str | None: 差异判断结果，None表示判断失败
|
||
"""
|
||
|
||
prompt = f"""请判断以下两个回答是否存在较大差异:
|
||
|
||
旧回答: {old_answer}
|
||
|
||
新回答: {new_answer}
|
||
|
||
主要是关键步骤、关键信息、或者关键主体的差异
|
||
请仅回答"存在较大差异"或"差异较小"。"""
|
||
|
||
try:
|
||
response = self.openai_llm.invoke(prompt)
|
||
return "无法判断,新老答案差异较大" if "存在较大差异" in response.content else "无法判断,新老答案基本相同"
|
||
except Exception as e:
|
||
return None
|
||
|
||
def process(self):
|
||
"""
|
||
处理所有问题并评判答案正确性
|
||
|
||
读取Excel文件中的问题和答案，进行评判，并将结果保存到输出Excel文件
|
||
"""
|
||
# 创建结果列表
|
||
results = []
|
||
|
||
# 读取Excel文件
|
||
for idx, row in tqdm(self.answer_excel.iterrows(), total=len(self.answer_excel), desc="处理问题"):
|
||
query = row["问题"]
|
||
old_answer = row["旧流程答案"]
|
||
new_answer = row["新流程答案"]
|
||
standard_answer = ""
|
||
|
||
try:
|
||
wiki_url = self.find_wiki_link(query)
|
||
if wiki_url and not pd.isna(wiki_url):
|
||
standard_answer = self.get_wiki_content(wiki_url)
|
||
except Exception as e:
|
||
print(f"处理问题 '{query}' 时发生错误: {str(e)}")
|
||
|
||
if standard_answer:
|
||
# 判断答案正确性
|
||
judge_result = self.judge_by_standard_answer(standard_answer, old_answer, new_answer)
|
||
else:
|
||
judge_result = self.judge_answer_diff(old_answer, new_answer)
|
||
|
||
if judge_result is None:
|
||
judge_result = ""
|
||
|
||
results.append({
|
||
"问题": query,
|
||
"旧流程答案": old_answer,
|
||
"新流程答案": new_answer,
|
||
"判断结果": judge_result
|
||
})
|
||
|
||
# 将结果转换为DataFrame并保存
|
||
results_df = pd.DataFrame(results)
|
||
results_df.to_excel(self.output_path, index=False)
|
||
print(f"处理完成，共处理 {len(results)} 条记录，结果已保存至 {self.output_path}")
|
||
|
||
# 测试函数
|
||
if __name__ == "__main__":
|
||
# 创建答案正确性评判工具实例
|
||
judge = AnswerCorrectnessJudge()
|
||
# 执行处理
|
||
judge.process() |