Update app.py

app.py

@@ -36,14 +36,15 @@ def chat_with_model(user_message, chat_history, max_new_tokens=60, temperature=0
 
     # Generate
     outputs = generator(
-        history_text,
-        max_new_tokens=int(max_new_tokens),
-        do_sample=True,
-        temperature=float(temperature),
-        top_p=float(top_p)
-    )
-
-    reply = outputs[0]["generated_text"][len(history_text):].strip()
+    history_text,
+    max_new_tokens=int(max_new_tokens),
+    do_sample=True,
+    temperature=float(temperature),
+    top_p=float(top_p)
+)
+
+    raw = outputs[0]["generated_text"][len(history_text):]
+    reply = clean_reply(raw)
 
     if "\n" in reply:
         reply = reply.split("\n")[0].strip()