Update app.py

app.py

@@ -367,14 +367,12 @@ def count_message_tokens(messages, model="gpt-3.5-turbo-0301"):
 
 def stream_notdiamond_response(response, model):
     """流式处理 notdiamond API 响应。"""
-    last_sent = ""
+    buffer = ""
     for chunk in response.iter_content(1024):
         if chunk:
             new_content = chunk.decode('utf-8')
-            to_send = new_content[len(last_sent):]
-            if to_send:
-                yield create_openai_chunk(to_send, model)
-            last_sent = new_content
+            buffer += new_content
+            yield create_openai_chunk(new_content, model)
     
     yield create_openai_chunk('', model, 'stop')