Update app.py

app.py

@@ -132,7 +132,6 @@ def messages():
         msg_id = str(uuid.uuid4())
         response_event = Event()
         response_text = []
-        output_tokens = 0  # 初始化 output_tokens
 
         if not stream:
             # 处理 stream 为 false 的情况
@@ -142,7 +141,6 @@ def messages():
         log_request(request.remote_addr, request.path, 200)
 
         def generate():
-            nonlocal output_tokens  # 使用 nonlocal 声明 output_tokens
             yield create_event("message_start", {
                 "type": "message_start",
                 "message": {
@@ -178,20 +176,12 @@ def messages():
                 sio.emit('perplexity_ask', (previous_messages, emit_data))
 
             def on_query_progress(data):
-                nonlocal response_text, output_tokens  # 使用 nonlocal 声明 response_text 和 output_tokens
+                nonlocal response_text
                 if 'text' in data:
                     text = json.loads(data['text'])
                     chunk = text['chunks'][-1] if text['chunks'] else None
                     if chunk:
                         response_text.append(chunk)
-                        output_tokens = calculate_tokens(''.join(response_text))  # 更新 output_tokens
-
-                        # 在 message_delta 事件中包含 output_tokens
-                        yield create_event("message_delta", {
-                            "type": "message_delta",
-                            "delta": {"stop_reason": None, "stop_sequence": None},
-                            "usage": {"input_tokens": input_tokens, "output_tokens": output_tokens},
-                        })
 
                 # 检查是否是最终响应
                 if data.get('final', False):
@@ -220,6 +210,13 @@ def messages():
                 
                 while not response_event.is_set():
                     sio.sleep(0.1)
+                    while response_text:
+                        chunk = response_text.pop(0)
+                        yield create_event("content_block_delta", {
+                            "type": "content_block_delta",
+                            "index": 0,
+                            "delta": {"type": "text_delta", "text": chunk},
+                        })
                 
             except Exception as e:
                 logging.error(f"Error during socket connection: {str(e)}")
@@ -232,6 +229,9 @@ def messages():
                 if sio.connected:
                     sio.disconnect()
 
+            # 动态计算输出的 token 数量
+            output_tokens = calculate_tokens(''.join(response_text))
+
             yield create_event("content_block_stop", {"type": "content_block_stop", "index": 0})
             yield create_event("message_delta", {
                 "type": "message_delta",