diff --git a/app/controller/add_knowledge_route.py b/app/controller/add_knowledge_route.py
index 0eb8ec5..e2c6c32 100644
--- a/app/controller/add_knowledge_route.py
+++ b/app/controller/add_knowledge_route.py
@@ -5,6 +5,8 @@ from typing import List
 
 from fastapi import UploadFile, File, Form
 from langchain_core.messages import HumanMessage
+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
@@ -21,26 +23,24 @@ def add_knowledge_route(app):
     return await milvus_service.async_search(param)
 
   @app.post("/knowledge/recall/stream")
-  async def knowledge_search(param: KnowledgeQueryParam):
+  async def knowledge_search(body: dict):
 
+    param = KnowledgeQueryParam(**body.get('input'))
     search_response = await milvus_service.async_search(param)
 
-    agent = create_react_agent(create_react_agent(
-      model=create_llm(),
-      prompt=f"""你需要根据如下内容来回答用户的问题：{search_response.answer}"""
-    ))
+    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():
       # 先把检索结果返回前端
-      yield f'data: {json.dumps({"type": "retrieve", "data": search_response}, ensure_ascii=False)}\n\n'
+      yield f'data: {json.dumps({"type": "retrieve", "data": search_response.model_dump()}, ensure_ascii=False)}\n\n'
 
-      async for chunk in agent.astream(
+      async for chunk in chain.astream(
         {"messages": [HumanMessage(content=param.question)]},
         stream_mode="messages"
       ):
-        # 流式回复
-        msg_chunk = chunk[0]
-        yield f'data: {json.dumps({"type": "messages", "data": msg_chunk.content}, ensure_ascii=False)}\n\n'
+        yield f'data: {json.dumps({"type": "messages", "data": chunk}, ensure_ascii=False)}\n\n'
 
     return StreamingResponse(generator_function(), media_type="text/event-stream")