''' Author: jackning 270580156@qq.com Date: 2024-08-29 14:49:54 LastEditors: jackning 270580156@qq.com LastEditTime: 2024-09-03 09:14:22 Description: bytedesk.com https://github.com/Bytedesk/bytedesk Please be aware of the BSL license restrictions before installing Bytedesk IM – selling, reselling, or hosting Bytedesk IM as a service is a breach of the terms and automatically terminates your rights under the license. 仅支持企业内部员工自用，严禁私自用于销售、二次销售或者部署SaaS方式销售 Business Source License 1.1: https://github.com/Bytedesk/bytedesk/blob/main/LICENSE contact: 270580156@qq.com 技术/商务联系：270580156@qq.com Copyright (c) 2024 by bytedesk.com, All Rights Reserved. ''' # import logging from typing import List from langchain_redis import RedisConfig, RedisVectorStore from langchain_huggingface import HuggingFaceEmbeddings from langchain_community.vectorstores.redis.filters import RedisFilter from langchain.docstore.document import Document from redisvl.query.filter import Tag from zhipuai import ZhipuAI from app.config import get_settings from app.redisPubsub import publishAnswerMessage, publishAnswerFinished zhipuAi = ZhipuAI(api_key=get_settings().ZHIPU_API_KEY) # 填写您自己的APIKey redis_index_schema = { "tag": [{"name": "kb_uid"}, {"name": "file_uid"}], } # https://python.langchain.com/v0.2/docs/integrations/vectorstores/redis/ class MyRedisVector: # embeddings: HuggingFaceEmbeddings vector_store: RedisVectorStore # def __init__(self): self.embeddings = HuggingFaceEmbeddings( model_name=get_settings().EMBEDDINGS_PATH) # config = RedisConfig( index_name=get_settings().REDIS_INDEX_NAME, key_prefix=get_settings().REDIS_KEY_PREFIX, redis_url=get_settings().REDIS_URL, redis_index_schema=redis_index_schema, metadata_schema=[ { "name": "kb_uid", "type": "tag", }, { "name": "file_uid", "type": "tag", } ] ) self.vector_store = RedisVectorStore(self.embeddings, config=config) return def add_docs(self, docs: List[Document]) -> List[str]: # 将docs中的文档存储到vector_store中 results = self.vector_store.add_documents(docs) logging.info(f"add_texts result: {results}") return results def delete_docs(self, docIds: List[str]) -> int: # 将docs从vector_store中删除 # 判断docIds是否为空，空则返回0 if not docIds: return 0 # 返回被删除的文档数量 return self.vector_store.index.drop_keys(docIds) # FIXME: RuntimeError: Index has not been created. Must be created before calling search def search_docs(self, kbUid: str, query: str) -> List[Document]: kb_filter = Tag("kb_uid") == kbUid results = self.vector_store.similarity_search( query, k=3, filter=kb_filter) # logging.info(f'search_store results {results}, {len(results)}') return results # 4.2 使用Retriever在向量库中搜索问题 # https://python.langchain.com/docs/integrations/vectorstores/redis#redis-as-retriever def search_as_retriever(self, kbUid: str, query: str) -> List[Document]: # query向量化？langchain文档demo中没有向量化 # kb_filter = RedisFilter.tag("kb_uid") == kbUid kb_filter = Tag("kb_uid") == kbUid # 有三种搜索算法 # 1 默认算法 retriever = self.vector_store.as_retriever( search_type="similarity", search_kwargs={ "k": 3, "filter": kb_filter } ) # 2 similarity_distance_threshold retriever which allows the user to specify the vector distance # retriever = self.vector_store.as_retriever( # search_type="similarity_distance_threshold", # search_kwargs={"k": 3, "distance_threshold": 0.1, # "filter": kb_filter}, # ) # 3 the similarity_score_threshold allows the user to define the minimum score for similar documents # retriever = self.vector_store.as_retriever( # search_type="similarity_score_threshold", # search_kwargs={"score_threshold": 0.9, # "k": 3, "filter": kb_filter}, # ) # results = retriever.get_relevant_documents(query) print(f'search_as_retriever results {results}, {len(results)}') return results # 5.Generation 将问题和搜索结果传给大模型，返回答案 def query_llm(self, messageUid: str, threadTopic: str, kbUid: str, question: str, search_results: List[Document]): # logging.info(f'query_llm search_results {search_results}, {len(search_results)}') # 拼接 # context = "\n".join(search_results) context = "\n".join([doc.page_content for doc in search_results]) # logging.info(f'query_llm context {context}') # 基于本地知识问答的提示词模 # TODO: write an english version # <问题 > {query} < /问题 > prompt_template = f''' ### Human: <指令>根据已知信息，简洁和专业的来回答问题。如果无法从中得到答案，请说 “根据已知信息无法回答该问题”，不允许在答案中添加编造成分，答案请使用跟问题中同样的语言。 <已知信息>{ context }### Assistant:''' print(f'prompt_template: { { prompt_template } }') # https://open.bigmodel.cn/dev/api#sdk_example response = zhipuAi.chat.completions.create( model="glm-4-flash", # 免费版 # prompt=[{"role": "user", "content": prompt_template}], messages=[ {"role": "system", "content": prompt_template}, {"role": "user", "content": question}, ], top_p=0.7, temperature=0.3, stream=True ) # counter = 0 for chunk in response: # ChatCompletionChunk(id='202408291708554ae5d8eac6e94630', choices=[Choice(delta=ChoiceDelta(content='凡', role='assistant', tool_calls=None), finish_reason=None, index=0)], created=1724922535, model='glm-4-air', usage=None, extra_json=None) # ChatCompletionChunk(id='202408291708554ae5d8eac6e94630', choices=[Choice(delta=ChoiceDelta(content='', role='assistant', tool_calls=None), finish_reason='stop', index=0)], created=1724922535, model='glm-4-air', usage=CompletionUsage(prompt_tokens=453, completion_tokens=34, total_tokens=487), extra_json=None) # print(f'chunk: { chunk }') # INFO: query_llm chunk: , stop, glm-4-air, CompletionUsage(prompt_tokens=453, completion_tokens=34, total_tokens=487) # counter += 1 id_to_publish = counter # answer = chunk.choices[0].delta.content model = chunk.model created = chunk.created finish_reason = chunk.choices[0].finish_reason logging.info( f'query_llm: {counter} {chunk.choices[0].delta.content}, {chunk.choices[0].finish_reason}, {chunk.model}, {chunk.created}, {chunk.usage}') if finish_reason == 'stop': promptTokens = chunk.usage.prompt_tokens completionTokens = chunk.usage.completion_tokens totalTokens = chunk.usage.total_tokens publishAnswerFinished( id=id_to_publish, uid=messageUid, threadTopic=threadTopic, kbUid=kbUid, question=question, answer=answer, model=model, created=created, promptTokens=promptTokens, completionTokens=completionTokens, totalTokens=totalTokens ) else: publishAnswerMessage( id=id_to_publish, uid=messageUid, threadTopic=threadTopic, kbUid=kbUid, question=question, answer=answer, model=model, created=created ) myredisVector = MyRedisVector() #