Update app.py

app.py

@@ -11,12 +11,12 @@ llm = Llama(
 )
 
 def chat(message, history):
-    # Enhanced system prompt for better responses
-    system_prompt = "You are a helpful, knowledgeable, and professional AI assistant. Provide detailed and thoughtful responses."
+    system_prompt = """Ti si YugoGPT, AI asistent koji prvenstveno komunicira na srpskom jeziku.
+    Tvoj zadatak je da daješ detaljne, istinite i korisne odgovore na srpskom jeziku."""
     
     full_prompt = f"""SYSTEM: {system_prompt}
 USER: {message}
-ASSISTANT: Let me provide a comprehensive response.
+ASSISTANT: Dozvolite mi da vam odgovorim na srpskom jeziku.
 """
     
     response = llm(
@@ -26,19 +26,24 @@ ASSISTANT: Let me provide a comprehensive response.
         top_p=0.95,
         repeat_penalty=1.2,
         top_k=40,
-        stop=["USER:", "\n\n"]
+        stop=["USER:", "\n\n"],
+        stream=True
     )
     
-    return response['choices'][0]['text']
+    partial_message = ""
+    for chunk in response:
+        if chunk and chunk['choices'][0]['text']:
+            partial_message += chunk['choices'][0]['text']
+            yield partial_message
 
 demo = gr.ChatInterface(
     fn=chat,
-    title="YugoGPT Professional Assistant",
-    description="I provide detailed and thoughtful responses to your questions.",
+    title="YugoGPT Asistent",
+    description="Postavite pitanje na srpskom ili bilo kom jeziku bivše Jugoslavije.",
     examples=[
-        "Explain quantum computing",
-        "What are the main principles of machine learning?",
-        "How does blockchain technology work?"
+        "Objasni kvantno računarstvo",
+        "Koje su glavne karakteristike veštačke inteligencije?",
+        "Kako funkcioniše blockchain tehnologija?"
     ]
 )