import json
import pathlib

# IMPORTANT: pandas must be imported BEFORE torch on Windows.
# They share an OpenMP/MKL DLL and whichever loads first wins; if torch
# loads first, importing pandas afterwards crashes the process silently
# (exit code 127). Keep pandas above torch.
import pandas as pd
import joblib
import torch

from src.data.preprocess import preprocess_data
from src.models.model import FraudMLP

# --- locate the artifacts/ folder at the project root ---
# this file is src/app/inference.py, so go up 2 levels to reach the root
ROOT = pathlib.Path(__file__).resolve().parents[2]
ARTIFACTS = ROOT / "artifacts"

# --- load the three saved artifacts ONCE when this module is imported ---
with open(ARTIFACTS / "metadata.json") as f:
    _metadata = json.load(f)

INPUT_DIM = _metadata["input_dim"]
FEATURE_COLUMNS = _metadata["feature_columns"]

_scaler = joblib.load(ARTIFACTS / "scaler.joblib")

_model = FraudMLP(input_dim=INPUT_DIM)
_model.load_state_dict(torch.load(ARTIFACTS / "model.pt", map_location="cpu"))
_model.eval()   # eval mode: BatchNorm/Dropout behave correctly for inference


def predict(transaction: dict) -> dict:
    """Score a single raw transaction for fraud.

    transaction: a dict with the raw fields (step, type, amount, the 4 balances)
    returns: {"fraud_probability": float, "is_fraud": int}
    """
    # 1. one-row DataFrame from the raw input
    df = pd.DataFrame([transaction])

    # 2. run the SAME preprocessing used in training
    processed = preprocess_data(df)

    # 3. guard: preprocess keeps only TRANSFER/CASH_OUT — anything else -> empty
    if processed.empty:
        raise ValueError("type must be TRANSFER or CASH_OUT")

    # 4. force the columns to match training exactly (adds missing type_* as 0,
    #    drops isFraud if present, orders columns like training)
    X = processed.reindex(columns=FEATURE_COLUMNS, fill_value=0)

    # 5. scale with the TRAINING scaler — transform only, never fit
    X_scaled = _scaler.transform(X)

    # 6. run the model
    tensor = torch.tensor(X_scaled, dtype=torch.float32)
    with torch.inference_mode():
        logit = _model(tensor)
        prob = torch.sigmoid(logit).item()
    return {
        "fraud_probability": prob,
        "is_fraud": int(prob > 0.5),
    }