Update app.py

app.py

@@ -478,22 +478,50 @@ with gr.Blocks(theme=gr.themes.Soft(), title="Chat with Sam") as demo:
     """)
     
     # Event handlers
+    def respond(message, chat_history, temperature, top_p, top_k, max_tokens):
+        """Handle message and generate response"""
+        # Add user message to history
+        chat_history.append([message, None])
+        
+        # Build conversation context from history (last 3 turns to save tokens)
+        conversation = ""
+        recent_history = chat_history[:-1][-3:] if len(chat_history) > 1 else []
+        
+        for user_msg, bot_msg in recent_history:
+            if bot_msg:  # Only include completed turns
+                conversation += f"User: {user_msg}\nSam: {bot_msg}\n"
+        
+        # Add current message
+        full_prompt = conversation + message if conversation else message
+        
+        # Generate with streaming
+        chat_history[-1][1] = ""
+        for response_chunk in generate_stream(
+            full_prompt, 
+            max_new_tokens=max_tokens,
+            temperature=temperature,
+            top_p=top_p,
+            top_k=top_k
+        ):
+            chat_history[-1][1] = response_chunk
+            yield chat_history
+    
     msg.submit(
-        chat_interface,
-        inputs=[msg, chatbot, temperature, top_p, top_k, max_tokens],
-        outputs=msg,
+        respond,
+        [msg, chatbot, temperature, top_p, top_k, max_tokens],
+        chatbot
     ).then(
-        lambda: gr.update(value=""),
+        lambda: gr.Textbox(value=""),
         None,
         msg
     )
     
     submit.click(
-        chat_interface,
-        inputs=[msg, chatbot, temperature, top_p, top_k, max_tokens],
-        outputs=msg,
+        respond,
+        [msg, chatbot, temperature, top_p, top_k, max_tokens],
+        chatbot
     ).then(
-        lambda: gr.update(value=""),
+        lambda: gr.Textbox(value=""),
         None,
         msg
     )