from __future__ import annotations import asyncio, json import uuid from typing import AsyncIterable, List, Optional, Literal from fastapi import Body, Request from fastapi.concurrency import run_in_threadpool from sse_starlette.sse import EventSourceResponse from langchain.callbacks import AsyncIteratorCallbackHandler from langchain.prompts.chat import ChatPromptTemplate from chatchat.settings import Settings from chatchat.server.agent.tools_factory.search_internet import search_engine from chatchat.server.api_server.api_schemas import OpenAIChatOutput from chatchat.server.chat.utils import History from chatchat.server.knowledge_base.kb_service.base import KBServiceFactory from chatchat.server.knowledge_base.kb_doc_api import search_docs, search_temp_docs from chatchat.server.knowledge_base.utils import format_reference from chatchat.server.utils import (wrap_done, get_ChatOpenAI, get_default_llm, BaseResponse, get_prompt_template, build_logger, check_embed_model, api_address ) logger = build_logger() async def kb_chat(query: str = Body(..., description="用户输入", examples=["你好"]), mode: Literal["local_kb", "temp_kb", "search_engine"] = Body("local_kb", description="知识来源"), kb_name: str = Body("", description="mode=local_kb时为知识库名称；temp_kb时为临时知识库ID，search_engine时为搜索引擎名称", examples=["samples"]), top_k: int = Body(Settings.kb_settings.VECTOR_SEARCH_TOP_K, description="匹配向量数"), score_threshold: float = Body( Settings.kb_settings.SCORE_THRESHOLD, description="知识库匹配相关度阈值，取值范围在0-1之间，SCORE越小，相关度越高，取到1相当于不筛选，建议设置在0.5左右", ge=0, le=2, ), history: List[History] = Body( [], description="历史对话", examples=[[ {"role": "user", "content": "我们来玩成语接龙，我先来，生龙活虎"}, {"role": "assistant", "content": "虎头虎脑"}]] ), stream: bool = Body(True, description="流式输出"), model: str = Body(get_default_llm(), description="LLM 模型名称。"), temperature: float = Body(Settings.model_settings.TEMPERATURE, description="LLM 采样温度", ge=0.0, le=2.0), max_tokens: Optional[int] = Body( Settings.model_settings.MAX_TOKENS, description="限制LLM生成Token数量，默认None代表模型最大值" ), prompt_name: str = Body( "default", description="使用的prompt模板名称(在prompt_settings.yaml中配置)" ), return_direct: bool = Body(False, description="直接返回检索结果，不送入 LLM"), request: Request = None, ): if mode == "local_kb": kb = KBServiceFactory.get_service_by_name(kb_name) if kb is None: return BaseResponse(code=404, msg=f"未找到知识库 {kb_name}") async def knowledge_base_chat_iterator() -> AsyncIterable[str]: try: nonlocal history, prompt_name, max_tokens history = [History.from_data(h) for h in history] if mode == "local_kb": kb = KBServiceFactory.get_service_by_name(kb_name) ok, msg = kb.check_embed_model() if not ok: raise ValueError(msg) docs = await run_in_threadpool(search_docs, query=query, knowledge_base_name=kb_name, top_k=top_k, score_threshold=score_threshold, file_name="", metadata={}) source_documents = format_reference(kb_name, docs, api_address(is_public=True)) elif mode == "temp_kb": ok, msg = check_embed_model() if not ok: raise ValueError(msg) docs = await run_in_threadpool(search_temp_docs, kb_name, query=query, top_k=top_k, score_threshold=score_threshold) source_documents = format_reference(kb_name, docs, api_address(is_public=True)) elif mode == "search_engine": result = await run_in_threadpool(search_engine, query, top_k, kb_name) docs = [x.dict() for x in result.get("docs", [])] source_documents = [f"""出处 [{i + 1}] [{d['metadata']['filename']}]({d['metadata']['source']}) \n\n{d['page_content']}\n\n""" for i,d in enumerate(docs)] else: docs = [] source_documents = [] # import rich # rich.print(dict( # mode=mode, # query=query, # knowledge_base_name=kb_name, # top_k=top_k, # score_threshold=score_threshold, # )) # rich.print(docs) if return_direct: yield OpenAIChatOutput( id=f"chat{uuid.uuid4()}", model=None, object="chat.completion", content="", role="assistant", finish_reason="stop", docs=source_documents, ) .model_dump_json() return callback = AsyncIteratorCallbackHandler() callbacks = [callback] # Enable langchain-chatchat to support langfuse import os langfuse_secret_key = os.environ.get('LANGFUSE_SECRET_KEY') langfuse_public_key = os.environ.get('LANGFUSE_PUBLIC_KEY') langfuse_host = os.environ.get('LANGFUSE_HOST') if langfuse_secret_key and langfuse_public_key and langfuse_host : from langfuse import Langfuse from langfuse.callback import CallbackHandler langfuse_handler = CallbackHandler() callbacks.append(langfuse_handler) if max_tokens in [None, 0]: max_tokens = Settings.model_settings.MAX_TOKENS llm = get_ChatOpenAI( model_name=model, temperature=temperature, max_tokens=max_tokens, callbacks=callbacks, ) # TODO： 视情况使用 API # # 加入reranker # if Settings.kb_settings.USE_RERANKER: # reranker_model_path = get_model_path(Settings.kb_settings.RERANKER_MODEL) # reranker_model = LangchainReranker(top_n=top_k, # device=embedding_device(), # max_length=Settings.kb_settings.RERANKER_MAX_LENGTH, # model_name_or_path=reranker_model_path # ) # print("-------------before rerank-----------------") # print(docs) # docs = reranker_model.compress_documents(documents=docs, # query=query) # print("------------after rerank------------------") # print(docs) context = "\n\n".join([doc["page_content"] for doc in docs]) if len(docs) == 0: # 如果没有找到相关文档，使用empty模板 prompt_name = "empty" prompt_template = get_prompt_template("rag", prompt_name) input_msg = History(role="user", content=prompt_template).to_msg_template(False) chat_prompt = ChatPromptTemplate.from_messages( [i.to_msg_template() for i in history] + [input_msg]) chain = chat_prompt | llm # Begin a task that runs in the background. task = asyncio.create_task(wrap_done( chain.ainvoke({"context": context, "question": query}), callback.done), ) if len(source_documents) == 0: # 没有找到相关文档 source_documents.append(f"未找到相关文档,该回答为大模型自身能力解答！") if stream: # yield documents first ret = OpenAIChatOutput( id=f"chat{uuid.uuid4()}", object="chat.completion.chunk", content="", role="assistant", model=model, docs=source_documents, ) yield ret.model_dump_json() async for token in callback.aiter(): ret = OpenAIChatOutput( id=f"chat{uuid.uuid4()}", object="chat.completion.chunk", content=token, role="assistant", model=model, ) yield ret.model_dump_json() else: answer = "" async for token in callback.aiter(): answer += token ret = OpenAIChatOutput( id=f"chat{uuid.uuid4()}", object="chat.completion", content=answer, role="assistant", model=model, ) yield ret.model_dump_json() await task except asyncio.exceptions.CancelledError: logger.warning("streaming progress has been interrupted by user.") return except Exception as e: logger.error(f"error in knowledge chat: {e}") yield {"data": json.dumps({"error": str(e)})} return if stream: return EventSourceResponse(knowledge_base_chat_iterator()) else: return await knowledge_base_chat_iterator().__anext__()