Update app.py

app.py

@@ -72,6 +72,7 @@ def chat_completions():
     def generate():
         nonlocal messages
         full_response = ""
+        is_first_response = True
         while True:
             conversation = "\n".join([f"{msg['role']}: {msg['content']}" for msg in messages])
             conversation += "\n请关注并回复user最近的消息并避免总结对话历史的回答"
@@ -92,6 +93,9 @@ def chat_completions():
                     full_response = data['text']
                     
                     if new_content:
+                        if is_first_response:
+                            yield f"data: {json.dumps(format_openai_response({'role': 'assistant'}))}\n\n"
+                            is_first_response = False
                         yield f"data: {json.dumps(format_openai_response(new_content))}\n\n"
                 
                 elif '"final":true' in event.data: