update stream

app.py

@@ -59,6 +59,10 @@ 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标记
+                in_think_block = False
+                found_final_answer = False
+                
                 for line in r.iter_lines():
                     if line:
                         line = line.decode('utf-8')
@@ -70,11 +74,31 @@ def respond(
                                     if 'content' in delta:
                                         content = delta['content']
                                         if content:
-                                            current_response += content
-                                            yield current_response
-                            except json.JSONDecodeError as e:
-                                print(f"[ERROR] JSON decode error: {e}")
+                                            # 将新内容添加到缓冲区
+                                            buffer += content
+                                            
+                                            # 检查缓冲区中是否包含完整的标记
+                                            if not in_think_block and '<think>' in buffer:
+                                                in_think_block = True
+                                                buffer = buffer.split('<think>')[-1]
+                                            
+                                            if in_think_block and '</think>' in buffer:
+                                                in_think_block = False
+                                                buffer = buffer.split('</think>')[-1]
+                                            
+                                            if not found_final_answer and '**Final Answer**' in buffer:
+                                                found_final_answer = True
+                                                buffer = buffer.split('**Final Answer**')[-1]
+                                            
+                                            # 如果不在think块内，就累积内容
+                                            if not in_think_block:
+                                                current_response += buffer
+                                                yield current_response
+                                                buffer = ""  # 清空缓冲区
+                                            
+                            except json.JSONDecodeError:
                                 continue
+                print(f"[INFO] final response: {current_response}")
             else:
                 print(f"[ERROR] Bad status code: {r.status_code}, response: {r.text}")
                 yield "Service temporarily unavailable"