Update app.py

app.py

@@ -65,10 +65,11 @@ if st.button("Send") and user_input:
     # ✅ Use a minimal prompt format (no system message)
     formatted_messages = [{"role": "user", "content": user_input}]
 
-    # ✅ Speed improvements: Reduce response length & force short answers
+    # ✅ Speed improvements: Increase max_tokens for full responses & ensure proper stopping
     response_data = st.session_state["model"].create_chat_completion(
         messages=formatted_messages, 
-        max_tokens=64, temperature=0.5, top_p=0.8,
+        max_tokens=128, temperature=0.5, top_p=0.8,
+        stop=["\n", "<|endoftext|>"],  # ✅ Ensures responses end properly
         stream=False  # ✅ No streaming for debugging
     )