app.py

import gradio as gr
from llama_cpp import Llama
from huggingface_hub import hf_hub_download

# 1. Téléchargement
print("⬇️ Downloading model...")
model_path = hf_hub_download(
    repo_id="XY26/dual-frame-llama-3",
    filename="meta-llama-3.1-8b-instruct.Q4_K_M.gguf"
)

# 2. Chargement du Moteur
print("⚙️ Loading engine...")
llm = Llama(
    model_path=model_path,
    n_ctx=4096,
    n_threads=2,
    verbose=False
)

def smart_response(message, history):
    try:
        messages = []
        
        # 1. System Prompt
        system_prompt = """You are a helpful AI assistant.
        - If asked for ADVICE/OPINION: Provide **Gain Frame** and **Loss Frame**.
        - If asked for FACTS/CHAT: Answer directly.
        Note: You are a small Quantized model, admit if you don't know."""
        
        messages.append({"role": "system", "content": system_prompt})

        # 2. Mémoire Sécurisée
        for turn in history:
            # On vérifie si c'est bien une liste avec au moins 2 éléments (User, Bot)
            if isinstance(turn, (list, tuple)) and len(turn) >= 2:
                user_msg = turn[0]
                bot_msg = turn[1]
                # On ajoute seulement si les messages ne sont pas vides
                if user_msg is not None and bot_msg is not None:
                    messages.append({"role": "user", "content": str(user_msg)})
                    messages.append({"role": "assistant", "content": str(bot_msg)})
        
        # 3. Question Actuelle
        messages.append({"role": "user", "content": message})

        # 4. Génération
        stream = llm.create_chat_completion(
            messages=messages,
            max_tokens=1024,
            stream=True,
            temperature=0.7
        )
        
        partial_message = ""
        for chunk in stream:
            delta = chunk['choices'][0]['delta']
            if 'content' in delta:
                token = delta['content']
                partial_message += token
                yield partial_message
                
    except Exception as e:
        # Si ça plante, on l'affiche proprement au lieu de casser l'interface
        print(f"❌ Error: {e}")
        yield f"⚠️ Oups, une erreur technique est survenue : {e}. Essayez de cliquer sur 'New Chat'."

# 3. Interface
demo = gr.ChatInterface(
    fn=smart_response,
    title="🤖 Smart Decision Architect (Safe Mode)",
    description="Ask factual questions or advice."
)

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