feat: 召回接口流式对话

app/controller/add_knowledge_route.py

@@ -4,10 +4,8 @@ from http.client import HTTPException
 from typing import List
 
 from fastapi import UploadFile, File, Form
-from langchain_core.messages import HumanMessage
+from langchain_core.messages import HumanMessage, SystemMessage
 from langchain_core.output_parsers import StrOutputParser
-from langchain_core.prompts import ChatPromptTemplate
-from langgraph.prebuilt import create_react_agent
 from starlette.requests import Request
 from starlette.responses import StreamingResponse
 
@@ -26,20 +24,19 @@ def add_knowledge_route(app):
   async def knowledge_search(body: dict):
 
     param = KnowledgeQueryParam(**body.get('input'))
-    search_response = await milvus_service.async_search(param)
-
-    chain = ChatPromptTemplate.from_template(
-      f"""你需要根据如下内容来回答用户的问题：{search_response.answer}"""
-    ) | create_llm() | StrOutputParser()
 
     async def generator_function():
+      search_response = await milvus_service.async_search(param)
+
+      chain = create_llm() | StrOutputParser()
+
       # 先把检索结果返回前端
       yield f'data: {json.dumps({"type": "retrieve", "data": search_response.model_dump()}, ensure_ascii=False)}\n\n'
 
-      async for chunk in chain.astream(
-        {"messages": [HumanMessage(content=param.question)]},
-        stream_mode="messages"
-      ):
+      async for chunk in chain.astream([
+        SystemMessage(content=f"""你需要根据如下内容来回答用户的问题：{search_response.answer}"""),
+        HumanMessage(content=param.question)
+      ]):
         yield f'data: {json.dumps({"type": "messages", "data": chunk}, ensure_ascii=False)}\n\n'
 
     return StreamingResponse(generator_function(), media_type="text/event-stream")