shorder prompt, removed history

a.py

@@ -2,7 +2,7 @@ from gradio_client import Client
 
 client = Client("wiklif/mixtral-api")
 result = client.predict(
-		prompt="Jak masz na  imie?",
+		prompt="Jak masz na imię?",
 		api_name="/chat"
 )
 print(result)

app.py

@@ -12,14 +12,11 @@ def generate_response(chat, kwargs):
         output += response.token.text
     return output
 
-def function(prompt, history):
-    chat = "<s>"
-    for user_prompt, bot_response in history:
-        chat += f"[INST] {user_prompt} [/INST] {bot_response}</s> "
-    chat += f"[INST] {prompt} [/INST]"
+def function(prompt):
+    chat = f"<s>[INST] {prompt} [/INST]</s>"
     kwargs = dict(
-        temperature=0.80,
-        max_new_tokens=2048,
+        temperature=0.5,
+        max_new_tokens=4096,
         top_p=0.95,
         repetition_penalty=1.0,
         do_sample=True,  # Upewnij się, że używasz próbkowania
@@ -28,9 +25,9 @@ def function(prompt, history):
 
     try:
         output = generate_response(chat, kwargs)
-        yield output
+        return output
     except:
-        yield ''
+        return ''
 
 interface = gr.ChatInterface(
     fn=function,