ouyangyouzhang/QueryRewrite
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

更新环境变量配置，调整模型名称获取方式，新增Dify API相关配置，删除无用的脚本文件，优化意图识别逻辑，添加LLM提取词条逻辑

Browse Source
This commit is contained in:
ouyangyouzhang
2025-07-16 14:24:50 +08:00
parent 5e164882a1
commit a934f2c398
28 changed files with 1834 additions and 1235 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rag2_0/intent_recognition/IntentRecognition.py
+105 -64
View File
@@ -39,10 +39,21 @@ from .ProfessionalNounVector import ProfessionalNounRetriever, AsyncProfessional
from rag2_0.tool.ModelTool import XinferenceReRankerModel, OpenAiLLM, SiliconFlowReRankerModel
class AsyncIntentRecognizer:
SOFT_WIKI_PATH = "data/wiki_data"
SOFT_NAMETOWIKI_MAP = {
"配网计价通D3软件": "配网计价通D3软件.txt",
"西藏计价通Z1软件": "西藏计价通Z1软件.txt",
"储能计价通C1软件": "储能计价通C1软件.txt",
"技改检修工程计价通T1软件": "技改检修工程计价通T1软件.txt",
"技改检修清单计价通T1软件": "技改检修清单计价通T1软件.txt",
"电力建设计价通软件": "电力建设计价通软件.txt",
"下载安装注册": "下载安装注册.txt",
}
"""
异步意图识别和问题改写类
"""
def __init__(self, api_key: str = None, base_url: str = None, model_name: str = "gpt-3.5-turbo", vector_index_dir: str = None):
def __init__(self):
"""
初始化异步意图识别器
@@ -52,51 +63,53 @@ class AsyncIntentRecognizer:
model_name: 要使用的模型名称
vector_index_dir: 向量索引目录，如果为None则使用默认目录
"""
api_key = os.getenv("OPENAI_API_KEY")
base_url = os.getenv("OPENAI_API_BASE")
model_name = os.getenv("MODEL_NAME", "gpt-3.5-turbo")
# 初始化LLM
llm_params = {
"temperature": 0.2, # 降低随机性，使结果更确定
"top_p": 0.7,
"model": model_name
"model": model_name,
"api_key": api_key,
"base_url": base_url
}
# 如果提供了API密钥，则使用提供的密钥
if api_key:
llm_params["api_key"] = api_key
# 如果提供了自定义URL，则使用提供的URL
if base_url:
llm_params["base_url"] = base_url
self._llm = OpenAiLLM(**llm_params)
llm_params["model"] = os.getenv("MINI_MODEL_NAME", "gpt-3.5-turbo")
self._llm_mini = OpenAiLLM(**llm_params)
# 加载suffix关键词
self._suffix_keywords = self._load_suffix_keywords()
# 加载软件词条名称库
self._soft_wiki_library = self._load_soft_wiki_library()
# 异步检索器将在create方法中初始化
self._noun_retriever = None
self._api_key = api_key
self._vector_index_dir = vector_index_dir
def _load_soft_wiki_library(self):
"""
加载软件wiki库
"""
SOFT_WIKI_LIBRARY = {}
for soft_name, wiki_file_name in self.SOFT_NAMETOWIKI_MAP.items():
with open(f"{self.SOFT_WIKI_PATH}/{wiki_file_name}", "r", encoding="utf-8") as f:
lines = f.readlines()
# 去除空行
lines = [line.strip() for line in lines if line.strip()]
SOFT_WIKI_LIBRARY[soft_name] = lines
return SOFT_WIKI_LIBRARY
@classmethod
async def create(cls, api_key: str = None, base_url: str = None, model_name: str = "gpt-3.5-turbo", vector_index_dir: str = None):
async def create(cls):
"""
异步工厂方法：创建并初始化异步意图识别器实例
Args:
api_key: OpenAI API密钥，如果为None则从环境变量获取
base_url: OpenAI API基础URL，如果为None则使用默认URL
model_name: 要使用的模型名称
vector_index_dir: 向量索引目录，如果为None则使用默认目录
Returns:
初始化完成的AsyncIntentRecognizer实例
"""
instance = cls(api_key, base_url, model_name, vector_index_dir)
instance = cls()
# 异步初始化名词检索器
instance._noun_retriever = await AsyncProfessionalNounRetriever.create(
api_key=api_key,
index_dir=vector_index_dir
)
instance._noun_retriever = await AsyncProfessionalNounRetriever.create()
return instance
def _load_suffix_keywords(self, filepath: str = None) -> List[str]:
@@ -402,11 +415,12 @@ class AsyncIntentRecognizer:
return f"通过博微软件助手查询软件锁信息，锁注册号为{lock_number}"
async def process_query_async(self, query: str, conversation_context: str = "",
async def process_query_async(self, query: str, conversation_context: Dict = None,
chat_history: List[Dict[str, str]] = None,
previous_slots: Dict[str, Any] = None,
use_jieba: bool = False,
enable_query_expansion: bool = False) -> Dict[str, Any]:
enable_query_expansion: bool = False,
cur_soft_name: str = "") -> Dict[str, Any]:
"""
异步处理用户问题的完整流程
@@ -417,7 +431,7 @@ class AsyncIntentRecognizer:
previous_slots: 历史槽位信息
use_jieba: 是否使用jieba分词辅助提取关键词
enable_query_expansion: 是否启用查询扩展
cur_soft_name: 当前查询的软件名称
Returns:
包含分类、关键词、改写和槽位填充结果的字典
"""
@@ -425,7 +439,8 @@ class AsyncIntentRecognizer:
chat_history = []
if previous_slots is None:
previous_slots = {}
if conversation_context is None:
conversation_context = {}
# 步骤: 并行执行提问扩展
query_expand_tasks = []
if enable_query_expansion:
@@ -437,9 +452,9 @@ class AsyncIntentRecognizer:
# 5.2: Follow Up Questions
asyncio.create_task(self._generate_follow_up_questions_async(query, chat_history, conversation_context)),
# 5.3: HyDE
# asyncio.create_task(self._generate_hypothetical_document_async(query, chat_history, conversation_context)),
# 5.3: 文档查询
asyncio.create_task(self._find_matching_software_docs_async(query, cur_soft_name, chat_history)),
# 5.4: 多问题查询
asyncio.create_task(self._generate_multi_questions_async(query, chat_history, conversation_context))
]
@@ -497,23 +512,22 @@ class AsyncIntentRecognizer:
# 收集结果
step_back_result = query_expand_results[0] if query_expand_results[0] else StepBackPrompt(original_query=query, can_use_back_prompt=False, step_back_query=[query])
follow_up_result = query_expand_results[1] if query_expand_results[1] else FollowUpQuestions(original_query=query, follow_up_query=query)
# hyde_result = query_expand_results[2] if query_expand_results[2] else HypotheticalDocument(original_query=query, hypothetical_answer="")
multi_questions_result = query_expand_results[2] if query_expand_results[2] else MultiQuestions(original_query=query, sub_questions=[query])
wiki_result = query_expand_results[2] if query_expand_results[2] else []
multi_questions_result = query_expand_results[3] if query_expand_results[3] else MultiQuestions(original_query=query, sub_questions=[query])
all_questions = multi_questions_result.sub_questions
all_questions.append(query)
all_questions.append(rewrite.rewrite)
all_questions.extend(step_back_result.step_back_query)
all_questions.append(follow_up_result.follow_up_query)
# all_questions.append(hyde_result.hypothetical_answer)
all_questions.extend(wiki_result)
all_questions = list(set(all_questions))
query_expand = {
"all": all_questions,
"step_back": step_back_result.model_dump(),
"follow_up": follow_up_result.model_dump(),
# "hyde": hyde_result.model_dump(),
"multi_questions": multi_questions_result.model_dump()
"multi_questions": multi_questions_result.model_dump(),
}
# 返回所有结果
@@ -721,45 +735,72 @@ class AsyncIntentRecognizer:
logging.error(f"异步后续问题生成失败: {e}", exc_info=True)
return FollowUpQuestions(original_query=query, follow_up_query=query)
async def _generate_hypothetical_document_async(self, query: str, chat_history: List[Dict[str, str]] = None, conversation_context: str = "") -> HypotheticalDocument:
async def _find_matching_software_docs_async(self, query: str, soft_name: str,
chat_history: List[Dict[str, str]] = None,
top_k: int = 3) -> List[str]:
"""
异步生成假设性文档
异步查找软件文档中与用户问题最匹配的几行内容
Args:
query: 用户原始问题
query: 用户问题
soft_name: 软件名称
chat_history: 历史对话记录
conversation_context: 会话背景信息
top_k: 返回的匹配行数，默认为3
Returns:
假设性文档结果
匹配的文档行列表
"""
if chat_history is None:
chat_history = []
# 检查软件名称是否在支持的列表中
if soft_name not in self.SOFT_NAMETOWIKI_MAP:
return []
# 获取软件文档内容
soft_docs = self._soft_wiki_library.get(soft_name, [])
if not soft_docs:
return []
soft_docs.extend(self._soft_wiki_library.get("下载安装注册", []))
# soft_docs=soft_docs[:50]
# 构建文档字符串，只包含行内容
soft_docs_str = "\n".join(f"{doc.strip()}" for i, doc in enumerate(soft_docs))
# 构建提示词，让LLM选择最匹配的行
prompt = f"""
{soft_docs_str}
================================
以上为软件功能操作、常见问题排查等功能，结合历史对话，请输出与当前提问最相关的1-3个功能名称，
使用Json格式输出，如下：
[{{"content": "行内容"}},...]
当前问题: {query}
历史对话: {json.dumps(chat_history, ensure_ascii=False)}
"""
hyde_start_time = time.time()
# 准备提示词
hyde_parser = PydanticOutputParser(pydantic_object=HypotheticalDocument)
formatted_prompt = hyde_prompt.format(
query=query,
chat_history=json.dumps(chat_history, ensure_ascii=False) if chat_history else "[]",
# conversation_context=conversation_context,
output_format=hyde_parser.get_format_instructions()
)
try:
# 异步调用LLM
response = await self._llm.invoke_async(formatted_prompt, False)
start_time = time.time()
response = await self._llm.invoke_async(prompt, False, response_format={"type": "json_object"})
end_time = time.time()
# 解析输出
response.content = response.content.strip()
clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
parsed_output = hyde_parser.parse(clean_output)
hyde_end_time = time.time()
hyde_time = hyde_end_time - hyde_start_time
logging.debug(f"异步假设性文档生成耗时统计 - 总耗时: {hyde_time:.2f}秒")
return parsed_output
# 解析JSON响应
try:
wiki_names = []
json_response = json.loads(response.content)
for match in json_response:
wiki_names.append(match["content"])
logging.debug(f"软件文档匹配耗时: {end_time - start_time:.2f}秒")
return wiki_names
except json.JSONDecodeError as e:
logging.error(f"解析JSON响应时出错: {e}")
return []
except Exception as e:
# 如果解析失败，返回空的假设性回答
logging.error(f"异步假设性文档生成失败: {e}", exc_info=True)
return HypotheticalDocument(original_query=query, hypothetical_answer="")
logging.error(f"查找匹配软件文档时出错: {e}", exc_info=True)
# 出错时返回空列表
return []
async def _generate_multi_questions_async(self, query: str, chat_history: List[Dict[str, str]] = None, conversation_context: str = "") -> MultiQuestions:
"""
异步生成多角度问题