app.py

import gradio as gr
from llama_cpp import Llama

llm = Llama(
    model_path="yugogpt-q4_k_s.gguf",
    n_ctx=4096,            # Doubled context length
    n_threads=8,           # Increased threads
    n_batch=1024,          # Increased batch size
    use_mlock=True,
    use_mmap=True,
    n_gpu_layers=0,        # Set this to higher number if GPU available
    verbose=False          # Reduced logging for better performance
)

def format_chat_history(history):
    formatted_history = ""
    for user_msg, assistant_msg in history:
        formatted_history += f"USER: {user_msg}\nA: {assistant_msg}\n"
    return formatted_history

def chat(message, history):
    system_prompt = """Ti si YugoGPT, profesionalni AI asistent koji daje precizne i korisne informacije.

PRAVILA:
- Ne izmišljam informacije
- Ako nemam trženu informaciju to jasno naglasim
- Dajem proverene jasne i konkretne informacije
- Koristim precizan srpski jezik
- Fokusiram se na činjenice
- Odgovaram direktno i efikasno
- Održavam profesionalan ton"""
    
    chat_history = format_chat_history(history)
    full_prompt = f"""SYSTEM: {system_prompt}

KONTEKST:
{chat_history}

USER: {message}
A: """
    
    response = llm(
        full_prompt,
        max_tokens=4096,    # Increased max tokens
        temperature=0.7,    # Keeping it precise
        top_p=0.1,
        repeat_penalty=1.2,
        top_k=20,
        stop=["USER:", "\n\n"],
        stream=True
    )
    
    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 Stručni Asistent",
    description="Profesionalni izvor informacija i stručne pomoći, PAŽNJA, ZNA DA LAŽE!!!",
    examples=[
        "Koji su osnovni principi relacionih baza podataka?",
        "Objasnite kako funkcioniše HTTP protokol",
        "Koje su glavne komponente računara i njihove funkcije?"
    ]
)

if __name__ == "__main__":
    demo.queue().launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False
    )