Update app.py

app.py

@@ -403,13 +403,15 @@ def handsome_chat_completions():
                 content_accumulated = ""
                 first_reasoning_chunk = True
                 
-                for chunk in response.iter_content(chunk_size=1024):
+                for chunk in response.iter_content(chunk_size=100000):
                     if chunk:
                         if first_chunk_time is None:
                             first_chunk_time = time.time()
                         full_response_content += chunk.decode("utf-8")
                         
                         for line in chunk.decode("utf-8").splitlines():
+                            # print(line)
+
                             if line.startswith("data:"):
                                 try:
                                     chunk_json = json.loads(line.lstrip("data: ").strip())
@@ -418,14 +420,17 @@ def handsome_chat_completions():
                                         
                                         if delta.get("reasoning_content") is not None:
                                             reasoning_chunk = delta["reasoning_content"]
+                                            reasoning_chunk = reasoning_chunk.replace('\n', '\n> ')
                                             if first_reasoning_chunk:
-                                                reasoning_chunk = "```Thinking\n" + reasoning_chunk
+                                                reasoning_chunk = "> " + reasoning_chunk
                                                 first_reasoning_chunk = False
                                             yield f"data: {json.dumps({'choices': [{'delta': {'content': reasoning_chunk}, 'index': 0}]})}\n\n"
                                             
                                         if delta.get("content") is not None:
                                             if not first_reasoning_chunk:
-                                                yield f"data: {json.dumps({'choices': [{'delta': {'content': '\n```\n'}, 'index': 0}]})}\n\n"
+                                                # yield f"data: {json.dumps({'choices': [{'delta': {'content': '\n```\n'}, 'index': 0}]})}\n\n"
+                                                # yield f"data: {json.dumps({'choices': [{'delta': {'content': '\n\n---\n\n### 结果输出\n'}, 'index': 0}]})}\n\n"
+                                                yield f"data: {json.dumps({'choices': [{'delta': {'content': '\n'}, 'index': 0}]})}\n\n"
                                                 first_reasoning_chunk = True
                                             yield f"data: {json.dumps({'choices': [{'delta': {'content': delta["content"]}, 'index': 0}]})}\n\n"
 
@@ -433,7 +438,6 @@ def handsome_chat_completions():
                                     logging.error(f"解析流式响应单行 JSON 失败: {e}, 行内容: {line}")
                                     continue
 
-
                 end_time = time.time()
                 first_token_time = (
                     first_chunk_time - start_time