update

app.py

@@ -54,8 +54,6 @@ def respond(
         with requests.post(API_URL, headers=headers, json=data, stream=True) as r:
             if r.status_code == 200:
                 current_response = ""
-                buffer = ""  # 用于累积可能被分割的内容
-                in_think_block = False
                 for line in r.iter_lines():
                     if line:
                         line = line.decode('utf-8')
@@ -67,35 +65,35 @@ def respond(
                                     if 'content' in delta:
                                         content = delta['content']
                                         if content:
-                                            buffer += content
+                                            current_response += content
                                             
-                                            if not in_think_block and '<think>' in buffer:
-                                                in_think_block = True
-                                                parts = buffer.split('<think>')
-                                                if len(parts) > 1:
-                                                    if parts[0]:
-                                                        current_response += parts[0]
-                                                    current_response += '```\n'
-                                                    buffer = parts[1]
-                                            
-                                            if in_think_block and ('</think>' in buffer or '**Final Answer**' in buffer):
-                                                in_think_block = False
-                                                current_response += buffer.split('</think>')[0] + '\n```'
-                                                if '</think>' in buffer:
-                                                    buffer = buffer.split('</think>')[1]
-                                                else:
-                                                    buffer = buffer.split('**Final Answer**')[1]
-                                            
-                                            if not in_think_block and buffer:
-                                                current_response += buffer
-                                                buffer = ""
-                                            
-                                            formatted_response = current_response
-                                            formatted_response = formatted_response.replace('<', '&lt;').replace('>', '&gt;')
-                                            formatted_response = formatted_response.replace('*', '\\*')
-                                            yield formatted_response
+                                            if len(current_response) > 16:
+                                                if '<think>' in current_response:
+                                                    current_response = current_response.replace('<think>', '```\n')
+                                                if '</think>' in current_response:
+                                                    current_response = current_response.replace('</think>', '\n```')
+                                                if '**Final Answer**' in current_response:
+                                                    current_response = current_response.replace('**Final Answer**', '')
+                                                
+                                                formatted_response = current_response[:-16]
+                                                
+                                                formatted_response = formatted_response.replace('<', '&lt;').replace('>', '&gt;')
+                                                formatted_response = formatted_response.replace('*', '\\*')
+                                                yield formatted_response
                             except json.JSONDecodeError:
                                 continue
+                if current_response:
+                    if '<think>' in current_response:
+                        current_response = current_response.replace('<think>', '```\n')
+                    if '</think>' in current_response:
+                        current_response = current_response.replace('</think>', '\n```')
+                    if '**Final Answer**' in current_response:
+                        current_response = current_response.replace('**Final Answer**', '')
+                    
+                    formatted_response = current_response
+                    formatted_response = formatted_response.replace('<', '&lt;').replace('>', '&gt;')
+                    formatted_response = formatted_response.replace('*', '\\*')
+                    yield formatted_response
             else:
                 print(f"[ERROR] Bad status code: {r.status_code}, response: {r.text}")
                 yield "Service temporarily unavailable"