Update app.py

app.py

@@ -48,15 +48,8 @@ def respond(message, max_tokens, temperature, top_p):
     thread = threading.Thread(target=run)
     thread.start()
 
-    response = ""
     for new_text in streamer:
-        response += new_text
-        # wrap model name and response in HTML
-        bubble = (
-            f"<div style='background:#eef2f5;padding:4px;border-radius:4px;font-size:small;max-width:fit-content;margin-bottom:4px;'>"
-            f"Model: {current_model_name}</div>" + response
-        )
-        yield bubble
+        yield new_text
 
 # User input handler
 def user(message, chat_history):
@@ -64,12 +57,18 @@ def user(message, chat_history):
     return "", chat_history
 
 # Bot response handler
-def bot(chatbot, max_tokens, temperature, top_p):
-    message = chatbot[-1][0]
-    response_generator = respond(message, max_tokens, temperature, top_p)
-    for response in response_generator:
-        chatbot[-1][1] = response
-        yield chatbot
+def bot(chat_history, max_tokens, temperature, top_p):
+    # Insert model name bubble and placeholder for output
+    chat_history.append([None, f"**Model:** {current_model_name}"])
+    chat_history.append([None, ""])
+    # Render model name immediately
+    yield chat_history
+
+    # Stream generation into the last bubble
+    message = chat_history[-3][0]  # Original user message
+    for chunk in respond(message, max_tokens, temperature, top_p):
+        chat_history[-1][1] += chunk
+        yield chat_history
 
 # Model selector handler
 def update_model(model_name):