main.py

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
import pandas as pd
import torch

# L'import marche parfaitement car le dossier 'model' a été cloné par Docker !
from model import Kronos, KronosTokenizer, KronosPredictor

app = FastAPI(title="Kronos Trading API")

print("📥 Chargement du modèle Kronos-small...")
try:
    tokenizer = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base")
    model = Kronos.from_pretrained("NeoQuasar/Kronos-small")
    
    # max_context est de 512 pour Kronos-small
    predictor = KronosPredictor(model, tokenizer, max_context=512)
    print("✅ Modèle chargé et prêt !")
except Exception as e:
    print(f"❌ ERREUR CRITIQUE : {str(e)}")
    predictor = None


class PredictionRequest(BaseModel):
    history_data: list
    lookback: int = 400
    pred_len: int = 24

@app.get("/")
def read_root():
    return {"status": "Kronos API is running 🚀", "model_loaded": predictor is not None}

# 💡 L'ASTUCE ANTI-ERREUR 404 : L'API répondra aux deux adresses !
@app.post("/predict")
@app.post("/predict/predict")
def predict(request: PredictionRequest):
    if predictor is None:
        raise HTTPException(status_code=500, detail="Le modèle n'a pas pu démarrer.")
        
    if not request.history_data:
        raise HTTPException(status_code=400, detail="Aucune donnée historique (OHLC) fournie.")
        
    df = pd.DataFrame(request.history_data)
    
    required_cols = ['open', 'high', 'low', 'close']
    if not all(col in df.columns for col in required_cols):
        raise HTTPException(status_code=400, detail=f"Colonnes manquantes. Requis : {required_cols}")

    if 'timestamps' not in df.columns:
        df['timestamps'] = pd.date_range(start='2024-01-01', periods=len(df), freq='H')
    else:
        df['timestamps'] = pd.to_datetime(df['timestamps'])

    actual_lookback = min(len(df), request.lookback, 512)
    x_df = df.tail(actual_lookback).reset_index(drop=True)
    x_timestamp = x_df['timestamps']
    
    last_time = x_timestamp.iloc[-1]
    
    # 👇 CORRECTION ICI : Conversion explicite en pd.Series 👇
    y_timestamp = pd.Series(pd.date_range(start=last_time, periods=request.pred_len + 1, freq='h')[1:])

    try:
        pred_df = predictor.predict(
            df=x_df,
            x_timestamp=x_timestamp,
            y_timestamp=y_timestamp,
            pred_len=request.pred_len,
            T=1.0,
            top_p=0.9,
            sample_count=1 
        )
        
        current_price = float(x_df['close'].iloc[-1])
        predicted_price = float(pred_df['close'].iloc[-1]) 
        expected_change = ((predicted_price - current_price) / current_price) * 100

        return {
            "status": "success",
            "current_price": current_price,
            "predicted_price_in_24h": predicted_price,
            "expected_change_percent": expected_change
        }
        
    except Exception as e:
        print(f"❌ Erreur lors de la prédiction : {str(e)}")
        raise HTTPException(status_code=500, detail=f"Erreur interne : {str(e)}")