上线THINK功能

routes/chat.py

@@ -26,11 +26,31 @@ agent_manager = init_global_sharded_agent_manager(
 )
 
 
-async def generate_stream_response(agent, messages, tool_response: bool, model: str):
+async def generate_stream_response(agent, messages, thought_list, tool_response: bool, model: str):
     """生成流式响应"""
     accumulated_content = ""
+
+
     chunk_id = 0
     try:
+
+        if len(thought_list)>0:
+            accumulated_content = get_content_from_messages(thought_list, tool_response=tool_response)
+            chunk_data = {
+                "id": f"chatcmpl-thought",
+                "object": "chat.completion.chunk",
+                "created": int(__import__('time').time()),
+                "model": model,
+                "choices": [{
+                    "index": 0,
+                    "delta": {
+                        "content": accumulated_content
+                    },
+                    "finish_reason": None
+                }]
+            }
+            yield f"data: {json.dumps(chunk_data, ensure_ascii=False)}\n\n"
+
         for response in agent.run(messages=messages):
             previous_content = accumulated_content
             accumulated_content = get_content_from_messages(response, tool_response=tool_response)
@@ -202,7 +222,8 @@ async def create_agent_and_generate_response(
 
             if all_checks:
                 # 将checks数组格式化为JSON字符串
-                guideline_analysis = json.dumps({"checks": all_checks}, ensure_ascii=False)
+                guideline_analysis = "\n".join([item["rationale"] for item in all_checks])
+                # guideline_analysis = json.dumps({"checks": all_checks}, ensure_ascii=False)
                 print(f"Merged guideline analysis result: {guideline_analysis}")
 
                 # 将分析结果添加到最后一个消息的内容中
@@ -226,17 +247,20 @@ async def create_agent_and_generate_response(
             user_identifier=user_identifier
         )
 
+    thought_list = []
+    if guideline_analysis != '':
+        thought_list = [{"role": "assistant","content": guideline_analysis}]
     # 根据stream参数决定返回流式还是非流式响应
     if stream:
         return StreamingResponse(
-            generate_stream_response(agent, messages, tool_response, model_name),
+            generate_stream_response(agent, messages, thought_list, tool_response, model_name),
             media_type="text/event-stream",
             headers={"Cache-Control": "no-cache", "Connection": "keep-alive"}
         )
     else:
         # 非流式响应
-        final_responses = agent.run_nonstream(messages)
+        agent_responses = agent.run_nonstream(messages)
-
+        final_responses = thought_list+agent_responses
         if final_responses and len(final_responses) > 0:
             # 使用 get_content_from_messages 处理响应，支持 tool_response 参数
             content = get_content_from_messages(final_responses, tool_response=tool_response)