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 = []
        
        # --- LE CERVEAU (AMÉLIORÉ) ---
        # On définit des règles beaucoup plus strictes pour forcer le format
        system_prompt = """You are a Decision Architect AI designed to eliminate bias by presenting two opposite perspectives.

        CORE INSTRUCTION:
        Analyze the user's input to determine if it is a "Simple Fact" or a "Subjective Topic".

        CATEGORY 1: SIMPLE FACTS (Math, Geography, History, Names)
        -> Example: "Capital of France?", "Who invented X?", "1+1?"
        -> ACTION: Answer directly, briefly, and normally.

        CATEGORY 2: SUBJECTIVE TOPICS (Decisions, Debates, Social Issues, Opinions, "Should I...")
        -> Example: "Nuclear energy?", "Remote work?", "Buy a warranty?", "Is X good?"
        -> ACTION: You MUST output exactly TWO distinct sections. Do not write an introduction or conclusion.
        
        FORMAT FOR CATEGORY 2:
        **The Gain Frame (Positive)**
        [Focus strictly on benefits, success, safety, opportunities, and why it is a GOOD idea.]

        **The Loss Frame (Negative)**
        [Focus strictly on costs, risks, dangers, failure, and why it is a BAD idea.]
        """
        
        messages.append({"role": "system", "content": system_prompt})

        # 2. Mémoire Sécurisée
        for turn in history:
            if isinstance(turn, (list, tuple)) and len(turn) >= 2:
                user_msg = turn[0]
                bot_msg = turn[1]
                if user_msg and bot_msg:
                    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
        # On augmente un peu la temperature (0.8) pour qu'il soit plus créatif dans ses arguments
        stream = llm.create_chat_completion(
            messages=messages,
            max_tokens=1024,
            stream=True,
            temperature=0.8 
        )
        
        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:
        print(f"❌ Error: {e}")
        yield f"⚠️ Erreur technique : {e}. Cliquez sur 'New Chat'."

# 3. Interface
demo = gr.ChatInterface(
    fn=smart_response,
    title="🤖 Smart Decision Architect (V2)",
    description="I differentiate between FACTS (Direct Answer) and DEBATES (Gain/Loss Frames)."
)

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