"""Central paths and constants for the Küchenpass-Agent project."""

from __future__ import annotations

import os
from pathlib import Path

PROJECT_ROOT = Path(__file__).resolve().parent.parent
DATA_DIR = PROJECT_ROOT / "data"
RAW_DIR = DATA_DIR / "raw"
PROCESSED_DIR = DATA_DIR / "processed"
SAMPLES_DIR = DATA_DIR / "samples"
MODELS_DIR = PROJECT_ROOT / "models"
DOCS_DIR = PROJECT_ROOT / "docs"
SCREENSHOTS_DIR = DOCS_DIR / "screenshots"

ML_MODEL_PATH = MODELS_DIR / "prep_time_xgb.joblib"
ML_PIPELINE_PATH = MODELS_DIR / "prep_time_pipeline.joblib"
ML_METRICS_PATH = MODELS_DIR / "ml_metrics.json"

CV_MODEL_PATH = MODELS_DIR / "food_classifier.pth"
CV_CLASSES_PATH = MODELS_DIR / "cv_classes.json"
CV_METRICS_PATH = MODELS_DIR / "cv_metrics.json"

# Geographic plausibility bounds for the Food Delivery dataset (India). Latitude/
# longitude outside these ranges are treated as noise and set to NaN during
# cleaning. Kept deliberately wide so valid rows are never dropped.
GEO_LAT_BOUNDS = (6.0, 38.0)
GEO_LON_BOUNDS = (68.0, 98.0)

# Restaurant-relevant subset of Food-101 (10 dishes that map to typical orders)
CV_TARGET_CLASSES: list[str] = [
    "pizza",
    "hamburger",
    "spaghetti_bolognese",
    "caesar_salad",
    "french_fries",
    "ramen",
    "steak",
    "sushi",
    "lasagna",
    "tiramisu",
]

STATIONS = ["grill", "pasta", "salad", "fryer", "pizza", "dessert", "cold"]

DISH_TO_STATION: dict[str, str] = {
    "pizza": "pizza",
    "hamburger": "grill",
    "burger": "grill",
    "steak": "grill",
    "spaghetti_bolognese": "pasta",
    "spaghetti": "pasta",
    "lasagna": "pasta",
    "pasta": "pasta",
    "ramen": "pasta",
    "caesar_salad": "salad",
    "salad": "salad",
    "french_fries": "fryer",
    "fries": "fryer",
    "sushi": "cold",
    "tiramisu": "dessert",
    "dessert": "dessert",
}

OPENAI_MODEL = os.getenv("OPENAI_MODEL", "gpt-4o-mini")

DECISION_CONFIDENCE_THRESHOLD = 0.6

# Maps NLP order dishes (which may carry style suffixes like "pizza_margherita"
# or synonyms like "cheeseburger") onto one of the 10 CV_TARGET_CLASSES so the
# pass-decision can compare the photographed dish with the order. Keys are
# matched first exactly, then as substrings (see canonical_cv_dish).
DISH_TO_CV_CLASS: dict[str, str] = {
    "pizza": "pizza",
    "margherita": "pizza",
    "hamburger": "hamburger",
    "burger": "hamburger",
    "cheeseburger": "hamburger",
    "spaghetti_bolognese": "spaghetti_bolognese",
    "spaghetti": "spaghetti_bolognese",
    "bolognese": "spaghetti_bolognese",
    "lasagna": "lasagna",
    "lasagne": "lasagna",
    "caesar_salad": "caesar_salad",
    "salad": "caesar_salad",
    "salat": "caesar_salad",
    "french_fries": "french_fries",
    "fries": "french_fries",
    "pommes": "french_fries",
    "ramen": "ramen",
    "steak": "steak",
    "sushi": "sushi",
    "tiramisu": "tiramisu",
}

# Per-class top-1 confidence thresholds for the pass decision. Derived from the
# threshold sweep in src/cv/evaluate.py (precision target 0.95 per class).
# Classes that are easily confused (lasagna<->pizza, spaghetti<->lasagna) need a
# higher bar before we send the plate to the guest. Missing classes fall back to
# DECISION_CONFIDENCE_THRESHOLD.
PER_CLASS_THRESHOLDS: dict[str, float] = {
    "lasagna": 0.85,
    "pizza": 0.75,
    "spaghetti_bolognese": 0.80,
    "hamburger": 0.75,
    "ramen": 0.70,
    "steak": 0.70,
    "caesar_salad": 0.60,
    "sushi": 0.60,
    "tiramisu": 0.55,
    "french_fries": 0.55,
}


for _d in (DATA_DIR, RAW_DIR, PROCESSED_DIR, SAMPLES_DIR, MODELS_DIR):
    _d.mkdir(parents=True, exist_ok=True)