Update app.py

app.py

@@ -25,14 +25,13 @@ client = OpenAI(
 
 # ✅ 回應函式 (流式 generator)
 def respond(message, history, system_message, max_tokens, temperature, top_p):
-    # history 是 list of dict: [{"role": "user"/"assistant", "content": "..."}]
     messages = [{"role": "system", "content": system_message}]
-    messages.extend(history)  # 直接加入舊對話
+    messages.extend(history)
     messages.append({"role": "user", "content": message})
 
     try:
         stream = client.chat.completions.create(
-            model="qwen3",   # ⚠️ 替換成你 llama.cpp 載入的模型 general.name
+            model="qwen3",
             messages=messages,
             max_tokens=max_tokens,
             temperature=temperature,
@@ -42,24 +41,20 @@ def respond(message, history, system_message, max_tokens, temperature, top_p):
         )
 
         output = ""
-        first_chunk_processed = False
+        skip_tokens = ['<|channel|>', 'analysis']
+        skipped = {token: False for token in skip_tokens}  # 追蹤每個 token 是否已忽略過
+
         for chunk in stream:
-            print("[DEBUG] chunk:", chunk)
             delta = chunk.choices[0].delta
-            if delta:              
-                # 安全取得 content
-                if delta.content:
-                    if not first_chunk_processed:
-                        first_chunk_processed = True
-                        if delta.content in ['<|channel|>', 'analysis']:
-                            continue
-                    output += delta.content
-                # 安全取得 reasoning_content，如果不存在就跳過 印出思維練
-                #reasoning = getattr(delta, "reasoning_content", None)
-                #if reasoning:
-                #    output += reasoning
-                # 每個 chunk 都 yield 最新內容
-                yield {"role": "assistant", "content": output}
+            if delta and delta.content:
+                content = delta.content.strip()
+                # 如果這個 token 是要跳過的，且還沒被跳過過
+                if content in skip_tokens and not skipped[content]:
+                    skipped[content] = True
+                    continue
+                output += delta.content  # 正常加入輸出
+
+            yield {"role": "assistant", "content": output}
 
     except Exception as e:
         print(f"[Error] {e}")