changed response

app.py

@@ -7,31 +7,31 @@ client = InferenceClient("google/gemma-3-27b-it")
 
 def respond(message, history):
     
-    messages = [{"role": "system", "content": "You are a British rude chatbot."}  # <to change the personality
+    messages = [{"role": "system", "content": "You are a rude British chatbot."}  # <to change the personality
                ]
     if history:
         messages.extend(history)
 
     messages.append({"role": "user", "content" : message})
     
-    response = client.chat_completion(
+    response = ""
+    for message in client.chat_completion(
         messages, 
-        max_tokens = 100, # to change the length
-        temperature = .9, #randomization
-        top_p = .7 
-    )
+        max_tokens = 500, # to change the length
+        stream = True, 
+        
+        # temperature = .9, #randomization
+        # top_p = .7 
+    ):
 
+        token = messages .choices[0].delta.content
+        response += token
+        yield response
     
     print(response)
     
-    return response ['choices'][0]['message']['content'].strip() 
+    #return response ['choices'][0]['message']['content'].strip() 
     
-def echo(message, history):
-    choices = ["yes", "ummmmm no no no", "sure", ".....I think you need to rethink your choices", "you deserve a sweet treat", "I don't think so"]
-    #use random to select one of these choices
-    random_message = random.choice(choices)
-    return random_message
-
 chatbot = gr.ChatInterface(respond, type = "messages", theme = "shivi/calm_seafoam")
 
 chatbot.launch(debug=True)