app.py

from flask import Flask, render_template, request, jsonify, current_app
import pandas as pd
import numpy as np
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics.pairwise import cosine_similarity
import os
import logging

# --- Logging Configuration ---
# Ensure logging is configured before any loggers are potentially used by imported modules
logging.basicConfig(level=logging.INFO, format='%(asctime)s %(levelname)s: %(message)s [in %(pathname)s:%(lineno)d]')
logger = logging.getLogger(__name__)

app = Flask(__name__)

# --- Global Variables ---
DF = None
ALL_TOPPINGS = []
FEATURE_DF = None
SCALER = None # Will be initialized in preprocess_data
NUMERICAL_COLS = ['Price', 'Slices', 'Rating', 'Spice_Level', 'Preparation_Time', 'Calories']
CATEGORICAL_FEATURES = [
    'Serving_Size', 'Popular_Group', 'Dietary_Category',
    'Sauce_Type', 'Cheese_Amount', 'Restaurant_Chain',
    'Seasonal_Availability', 'Bread_Type'
]
CRUST_TYPE_COL = None
DEFAULT_IMAGE_URL = 'https://images.dominos.co.in/new_margherita_2502.jpg'


def preprocess_data(df_path='pizza.csv'):
    global DF, ALL_TOPPINGS, FEATURE_DF, SCALER, CATEGORICAL_FEATURES, CRUST_TYPE_COL
    logger.info(f"Attempting to preprocess data from relative path: {df_path}")

    # Construct absolute path for the CSV file
    # This is crucial for environments like Docker where working directory might differ
    base_dir = os.path.dirname(os.path.abspath(__file__)) # Directory of the current script (app.py)
    absolute_df_path = os.path.join(base_dir, df_path)
    logger.info(f"Absolute path for CSV: {absolute_df_path}")

    if not os.path.exists(absolute_df_path):
        logger.error(f"Dataset file '{absolute_df_path}' not found.")
        raise FileNotFoundError(f"Dataset file '{absolute_df_path}' not found. Ensure it's in the same directory as app.py.")

    DF = pd.read_csv(absolute_df_path)
    logger.info(f"Successfully loaded '{absolute_df_path}'. Original DataFrame shape: {DF.shape}")
    logger.info(f"Original DataFrame columns: {DF.columns.tolist()}")

    # Determine Crust Type Column
    potential_crust_cols = ['Crust_Type', 'Cr_Type']
    valid_crust_cols = [col for col in potential_crust_cols if col in DF.columns]
    if valid_crust_cols:
        valid_crust_cols.sort(key=lambda col: DF[col].isnull().sum()) # Prefer column with fewer NaNs
        CRUST_TYPE_COL = valid_crust_cols[0]
        logger.info(f"Using '{CRUST_TYPE_COL}' for crust type.")
        if CRUST_TYPE_COL not in CATEGORICAL_FEATURES:
            CATEGORICAL_FEATURES.append(CRUST_TYPE_COL)
        # Remove other potential crust columns if they were in CATEGORICAL_FEATURES
        for col in potential_crust_cols:
            if col != CRUST_TYPE_COL and col in CATEGORICAL_FEATURES:
                CATEGORICAL_FEATURES.remove(col)
    else:
        logger.warning("Crust type column (Crust_Type or Cr_Type) not found. Crust type will not be used.")
        CRUST_TYPE_COL = None

    # Fill NaN for text-based categorical columns and other text fields
    text_cols_to_fill = list(set(CATEGORICAL_FEATURES + ['Toppings', 'Description', 'Allergens', 'Image_Url', 'Pizza_Name']))
    for col in text_cols_to_fill:
        if col and col in DF.columns: # Ensure col is not None (e.g. if CRUST_TYPE_COL is None)
            DF[col] = DF[col].fillna('')
    logger.info("Filled NaNs in text-based categorical columns with empty strings.")

    # Fill NaN for numerical columns from the CSV
    numerical_cols_in_df = ['Price_Rs', 'Slices', 'Rating', 'Rating_Count', 'Preparation_Time_min', 'Calories_per_Slice']
    for col in numerical_cols_in_df:
        if col in DF.columns:
            if pd.api.types.is_numeric_dtype(DF[col]):
                median_val = DF[col].median()
                DF[col] = DF[col].fillna(median_val)
                logger.info(f"Filled NaNs in numerical column '{col}' with its median ({median_val}).")
            else:
                # Attempt to convert to numeric, then fill with median or 0
                numeric_series = pd.to_numeric(DF[col], errors='coerce')
                median_val = 0
                if not numeric_series.isnull().all():
                    median_val = numeric_series.median()
                DF[col] = numeric_series.fillna(median_val)
                logger.warning(f"Column '{col}' was not purely numeric. Converted to numeric, filled NaNs with median/0 ({median_val}).")
        else:
            logger.warning(f"Expected numerical column '{col}' not found in DataFrame. It will be missing from features if not handled.")


    if 'Rating_Count' in DF.columns:
        DF['Rating_Count'] = DF['Rating_Count'].fillna(0).astype(int)

    # Process Toppings
    if 'Toppings' in DF.columns:
        DF['Toppings_list_internal'] = DF['Toppings'].astype(str).str.split(r';\s*') # Use raw string for regex
        DF['Toppings_list_internal'] = DF['Toppings_list_internal'].apply(
            lambda x: [t.strip() for t in x if isinstance(t, str) and t.strip()]) # Filter out empty strings after split
        current_all_toppings = set()
        for toppings_list in DF['Toppings_list_internal'].dropna():
            current_all_toppings.update(t for t in toppings_list if t) # Ensure t is not empty
        ALL_TOPPINGS = sorted(list(current_all_toppings))
        logger.info(f"Found {len(ALL_TOPPINGS)} unique toppings. Example: {ALL_TOPPINGS[:5] if ALL_TOPPINGS else 'None'}")
    else:
        logger.warning("'Toppings' column not found. Topping features will be empty.")
        DF['Toppings_list_internal'] = pd.Series([[] for _ in range(len(DF))]) # Empty list for all rows
        ALL_TOPPINGS = []


    # --- Feature Engineering ---
    feature_data = {}
    num_feature_map = {
        'Price': 'Price_Rs', 'Slices': 'Slices', 'Rating': 'Rating',
        'Preparation_Time': 'Preparation_Time_min', 'Calories': 'Calories_per_Slice'
    }
    for feature_col, df_col in num_feature_map.items():
        if df_col in DF.columns:
            feature_data[feature_col] = DF[df_col].copy()
        else:
            logger.warning(f"Numerical source column '{df_col}' for feature '{feature_col}' not found. Filling with zeros.")
            feature_data[feature_col] = pd.Series([0.0] * len(DF)) # Ensure float for consistency

    # Spice Level Feature (Numerical)
    if 'Spice_Level' in DF.columns:
        DF['Spice_Level'] = DF['Spice_Level'].fillna('Mild') # Default for NaNs
        spice_map = {'Mild': 1, 'Medium': 2, 'Hot': 3}
        feature_data['Spice_Level'] = DF['Spice_Level'].map(spice_map).fillna(1.0) # Ensure float
    else:
        logger.warning("'Spice_Level' column not found. Filling 'Spice_Level' feature with default (1.0).")
        feature_data['Spice_Level'] = pd.Series([1.0] * len(DF)) # Default if column is missing

    # One-Hot Encode Categorical Features
    for feature_cat_col in CATEGORICAL_FEATURES:
        if feature_cat_col and feature_cat_col in DF.columns: # Check if col_name is not None and exists
            # Ensure the column is treated as string to avoid issues with mixed types in unique()
            DF[feature_cat_col] = DF[feature_cat_col].astype(str)
            for value in DF[feature_cat_col].unique():
                if pd.notnull(value) and value.strip() != '': # Check for non-null and non-empty string values
                    feature_data[f"{feature_cat_col}_{value}"] = (DF[feature_cat_col] == value).astype(int)
        elif feature_cat_col: # Log warning only if feature_cat_col was defined
             logger.warning(f"Categorical source column '{feature_cat_col}' for one-hot encoding not found in DataFrame.")

    # Topping Features (One-Hot Encoded)
    for topping in ALL_TOPPINGS:
        if topping: # Ensure topping string is not empty
            feature_data[f"Topping_{topping}"] = DF['Toppings_list_internal'].apply(
                lambda x: 1 if topping in x else 0
            )

    FEATURE_DF = pd.DataFrame(feature_data)
    logger.info(f"FEATURE_DF created. Shape: {FEATURE_DF.shape}. Columns: {FEATURE_DF.columns.tolist()[:10]}...") # Log first 10 cols

    # Ensure all NUMERICAL_COLS exist in FEATURE_DF and fill NaNs
    for col in NUMERICAL_COLS:
        if col not in FEATURE_DF.columns:
            logger.warning(f"Numerical column '{col}' is missing from FEATURE_DF after construction. Adding as zeros.")
            FEATURE_DF[col] = 0.0 # Ensure float
        if FEATURE_DF[col].isnull().any():
            mean_val = FEATURE_DF[col].mean()
            fill_val = mean_val if pd.notna(mean_val) else 0.0
            logger.info(f"Filling NaNs in numerical feature column '{col}' with {fill_val}.")
            FEATURE_DF[col] = FEATURE_DF[col].fillna(fill_val)

    # Scale Numerical Features
    SCALER = MinMaxScaler() # Initialize scaler
    if not FEATURE_DF.empty and all(col in FEATURE_DF.columns for col in NUMERICAL_COLS):
        try:
            FEATURE_DF[NUMERICAL_COLS] = SCALER.fit_transform(FEATURE_DF[NUMERICAL_COLS])
            logger.info(f"Numerical columns ({NUMERICAL_COLS}) scaled. FEATURE_DF shape: {FEATURE_DF.shape}")
        except Exception as e:
            logger.error(f"Error during scaling of numerical columns: {e}. FEATURE_DF might be problematic.")
            # Fallback: Keep numerical columns unscaled if scaling fails, or handle as needed
    elif FEATURE_DF.empty:
        logger.error("FEATURE_DF is empty before scaling. Scaling skipped. This will likely cause issues.")
    else:
        missing_cols = [col for col in NUMERICAL_COLS if col not in FEATURE_DF.columns]
        logger.error(f"Not all numerical columns ({NUMERICAL_COLS}) found in FEATURE_DF for scaling. Missing: {missing_cols}. Scaling skipped.")

    logger.info(f"Preprocessing done. DF is None: {DF is None}, FEATURE_DF is None: {FEATURE_DF is None}, SCALER is None: {SCALER is None}")
    if FEATURE_DF is not None:
        logger.info(f"Final FEATURE_DF shape: {FEATURE_DF.shape}")
    if DF is not None:
        logger.info(f"Final DF shape: {DF.shape}")


@app.route('/')
def index_route():
    global DF, ALL_TOPPINGS, CATEGORICAL_FEATURES, CRUST_TYPE_COL, FEATURE_DF, DEFAULT_IMAGE_URL
    # Critical check at the beginning of the route
    if DF is None:
        current_app.logger.error("DF is None when trying to serve '/'. Data preprocessing might have failed or not run.")
        return "Error: Pizza data (DF) not loaded. Please check server logs.", 500
    if FEATURE_DF is None: # Also check FEATURE_DF as it's derived
        current_app.logger.error("FEATURE_DF is None when trying to serve '/'. Data preprocessing might have failed.")
        return "Error: Pizza feature data (FEATURE_DF) not loaded. Please check server logs.", 500

    filter_options = {}
    # Ensure 'Spice_Level' is included for filter options if it exists in DF
    cols_for_filters_set = set(cat_col for cat_col in CATEGORICAL_FEATURES if cat_col and cat_col in DF.columns) # Filter out None or non-existent
    if 'Spice_Level' in DF.columns:
        cols_for_filters_set.add('Spice_Level')
    # CRUST_TYPE_COL is already in CATEGORICAL_FEATURES if found

    for col_name in list(cols_for_filters_set):
        # key_name for JS should be consistent (lowercase, no underscores)
        key_name = col_name.lower().replace('_', '')
        # No special handling for spicelevel or crusttype here, it's naturally handled by the line above.

        unique_values = sorted([v for v in DF[col_name].astype(str).dropna().unique() if v.strip() != ''])
        if unique_values: # Only add if there are actual values
            filter_options[key_name] = unique_values

    # Prepare default recommendations (e.g., top-rated)
    # Make sure 'Rating' column exists
    if 'Rating' in DF.columns:
        default_recommendations_df = DF.sort_values('Rating', ascending=False).copy()
    else:
        logger.warning("'Rating' column not found in DF. Cannot sort for default recommendations. Using unsorted DF.")
        default_recommendations_df = DF.copy() # Fallback to unsorted

    default_recs_list = []
    frontend_keys = [
        'id', 'name', 'toppings', 'price', 'slices', 'serving_size', 'rating', 'rating_count',
        'description', 'popular_group', 'dietary_category', 'spice_level', 'sauce_type',
        'cheese_amount', 'calories', 'allergens', 'prep_time', 'restaurant', 'seasonal',
        'bread_type', 'image_url', 'crust_type'
    ]
    df_to_frontend_map = {
        'id': None, 'name': 'Pizza_Name', 'toppings': 'Toppings', 'price': 'Price_Rs', 'slices': 'Slices',
        'serving_size': 'Serving_Size', 'rating': 'Rating', 'rating_count': 'Rating_Count',
        'description': 'Description', 'popular_group': 'Popular_Group',
        'dietary_category': 'Dietary_Category', 'spice_level': 'Spice_Level',
        'sauce_type': 'Sauce_Type', 'cheese_amount': 'Cheese_Amount',
        'calories': 'Calories_per_Slice', 'allergens': 'Allergens',
        'prep_time': 'Preparation_Time_min', 'restaurant': 'Restaurant_Chain',
        'seasonal': 'Seasonal_Availability', 'bread_type': 'Bread_Type',
        'image_url': 'Image_Url', 'crust_type': CRUST_TYPE_COL # Uses the determined CRUST_TYPE_COL
    }

    for original_idx, pizza_row in default_recommendations_df.iterrows():
        rec_item = {}
        for key in frontend_keys:
            df_col = df_to_frontend_map.get(key)
            if key == 'id':
                rec_item[key] = int(original_idx) # Pizza ID is its original index in DF
            elif df_col and df_col in pizza_row: # df_col can be None for 'id' or if CRUST_TYPE_COL is None
                value = pizza_row[df_col]
                # Type conversions for JSON serializability
                if isinstance(value, np.integer): value = int(value)
                elif isinstance(value, np.floating): value = float(value)
                elif isinstance(value, np.ndarray): value = value.tolist()
                rec_item[key] = "" if pd.isna(value) else value
            elif key == 'crust_type' and not CRUST_TYPE_COL : # If CRUST_TYPE_COL was not found
                 rec_item[key] = "N/A"
            else:
                rec_item[key] = "" # Default for missing fields

        rec_item['rating_count'] = int(rec_item.get('rating_count', 0) or 0) # Ensure int
        rec_item['image_url'] = rec_item.get('image_url') if rec_item.get('image_url') else DEFAULT_IMAGE_URL

        # Final pass to convert any remaining numpy generic types
        for k_final, v_final in rec_item.items():
            if isinstance(v_final, np.generic): rec_item[k_final] = v_final.item()
        default_recs_list.append(rec_item)

    current_app.logger.info(f"Serving {len(default_recs_list)} pizzas for initial display.")
    current_app.logger.info(f"Filter options for template: {filter_options}")
    current_app.logger.info(f"ALL_TOPPINGS for template: {ALL_TOPPINGS[:5] if ALL_TOPPINGS else 'None'}")


    return render_template('index.html',
                           toppings=ALL_TOPPINGS,
                           filter_options=filter_options,
                           default_recommendations=default_recs_list,
                           default_image_url=DEFAULT_IMAGE_URL)


def get_recommendations(preferences):
    global DF, FEATURE_DF, SCALER, CRUST_TYPE_COL, DEFAULT_IMAGE_URL

    if DF is None or FEATURE_DF is None or SCALER is None:
        current_app.logger.error("Data not fully initialized (DF, FEATURE_DF, or SCALER is None) for get_recommendations.")
        return []

    current_indices = DF.index.to_list()
    current_app.logger.info(f"Starting with {len(current_indices)} pizzas before filtering. Preferences: {preferences}")

    # --- Hard Filters ---
    # 1. Toppings
    if 'toppings' in preferences and preferences['toppings'] and 'Toppings_list_internal' in DF.columns:
        selected_toppings = set(preferences['toppings'])
        if selected_toppings: # Ensure not an empty list that would select nothing
            topping_mask = DF.loc[current_indices, 'Toppings_list_internal'].apply(
                lambda x_toppings: isinstance(x_toppings, list) and any(t in selected_toppings for t in x_toppings)
            )
            current_indices = DF.loc[current_indices][topping_mask].index.to_list()
            current_app.logger.info(f"After toppings filter: {len(current_indices)} pizzas remaining")
            if not current_indices: return []

    # 2. Max Price
    if 'price_range' in preferences and preferences['price_range'] and 'Price_Rs' in DF.columns:
        try:
            min_price = float(preferences['price_range'][0])
            max_price = float(preferences['price_range'][1])
            price_mask = (DF.loc[current_indices, 'Price_Rs'] >= min_price) & \
                         (DF.loc[current_indices, 'Price_Rs'] <= max_price)
            current_indices = DF.loc[current_indices][price_mask].index.to_list()
            current_app.logger.info(f"After price filter ({min_price}-{max_price}): {len(current_indices)} pizzas")
            if not current_indices: return []
        except (TypeError, ValueError, IndexError) as e:
            current_app.logger.warning(f"Invalid price_range preference: {preferences['price_range']}. Error: {e}")


    # 3. Number of Slices (Min Slices)
    if 'slices' in preferences and preferences['slices'] is not None and 'Slices' in DF.columns:
        try:
            min_slices = int(preferences['slices'])
            slices_mask = DF.loc[current_indices, 'Slices'] >= min_slices
            current_indices = DF.loc[current_indices][slices_mask].index.to_list()
            current_app.logger.info(f"After slices filter (>= {min_slices}): {len(current_indices)} pizzas")
            if not current_indices: return []
        except ValueError:
            current_app.logger.warning(f"Invalid value for slices: {preferences['slices']}")

    # 4. Minimum Rating
    if 'rating' in preferences and preferences['rating'] is not None and 'Rating' in DF.columns:
        try:
            min_rating = float(preferences['rating'])
            rating_mask = DF.loc[current_indices, 'Rating'] >= min_rating
            current_indices = DF.loc[current_indices][rating_mask].index.to_list()
            current_app.logger.info(f"After rating filter (>= {min_rating}): {len(current_indices)} pizzas")
            if not current_indices: return []
        except ValueError:
            current_app.logger.warning(f"Invalid value for rating: {preferences['rating']}")

    # 5. Max Preparation Time
    if 'prep_time' in preferences and preferences['prep_time'] is not None and 'Preparation_Time_min' in DF.columns:
        try:
            max_prep_time = int(str(preferences['prep_time']).lower().replace("min", "").strip())
            prep_mask = DF.loc[current_indices, 'Preparation_Time_min'] <= max_prep_time
            current_indices = DF.loc[current_indices][prep_mask].index.to_list()
            current_app.logger.info(f"After prep time filter (<= {max_prep_time}): {len(current_indices)} pizzas")
            if not current_indices: return []
        except ValueError:
            current_app.logger.warning(f"Could not parse prep_time value: {preferences['prep_time']}")

    # 6. Categorical Filters (Multi-select OR logic)
    # JS keys: servingsize, populargroup, dietarycategory, spicelevel, saucetype, etc.
    categorical_pref_map = {
        "servingsize": "Serving_Size", "populargroup": "Popular_Group",
        "dietarycategory": "Dietary_Category", "spicelevel": "Spice_Level",
        "saucetype": "Sauce_Type", "cheeseamount": "Cheese_Amount",
        "restaurantchain": "Restaurant_Chain", "seasonalavailability": "Seasonal_Availability",
        "breadtype": "Bread_Type", "crusttype": CRUST_TYPE_COL
    }
    for pref_key, df_col_name in categorical_pref_map.items():
        if df_col_name and pref_key in preferences and preferences[pref_key]: # Ensure df_col_name is not None
            pref_value_list = preferences[pref_key] # Expected to be a list from JS
            if isinstance(pref_value_list, list) and pref_value_list: # If list is not empty
                if df_col_name in DF.columns:
                    cat_mask = DF.loc[current_indices, df_col_name].isin(pref_value_list)
                    current_indices = DF.loc[current_indices][cat_mask].index.to_list()
                    current_app.logger.info(f"After {pref_key} filter (isin {pref_value_list}): {len(current_indices)} pizzas")
                    if not current_indices: return []
                else:
                    current_app.logger.warning(f"Column '{df_col_name}' for preference '{pref_key}' not found in DF. Filter skipped.")
            # If pref_value_list is empty, it means "Any" for this category, so no filtering.

    if not current_indices:
        current_app.logger.info("No pizzas match all hard filter criteria.")
        return []

    # --- Similarity Scoring Part ---
    # Filter FEATURE_DF to only include pizzas remaining after hard filters
    valid_indices_for_feature_df = FEATURE_DF.index.intersection(current_indices)
    if valid_indices_for_feature_df.empty:
        current_app.logger.info("No valid indices remain for FEATURE_DF after hard filters.")
        return []

    filtered_feature_df = FEATURE_DF.loc[valid_indices_for_feature_df]
    if filtered_feature_df.empty: # Should not happen if valid_indices_for_feature_df is not empty
        current_app.logger.warning("Filtered FEATURE_DF is empty. This is unexpected.")
        return []

    # Create User Preference Vector (aligned with FEATURE_DF columns)
    user_vector = pd.Series(0.0, index=FEATURE_DF.columns) # Initialize with 0.0 for float consistency

    # 1. Toppings in User Vector
    if 'toppings' in preferences and preferences['toppings']:
        for topping in preferences['toppings']:
            col_name = f"Topping_{topping}"
            if col_name in user_vector.index:
                user_vector[col_name] = 1.0

    # 2. Categorical Preferences (One-Hot) in User Vector
    # js_to_df_key_map_for_vector is same as categorical_pref_map but df_col_name is for one-hot prefix
    for pref_key, df_col_prefix in categorical_pref_map.items():
        if df_col_prefix and pref_key in preferences and preferences[pref_key]: # df_col_prefix can be None for CRUST_TYPE_COL
            selected_values = preferences[pref_key] # This is a list
            for val_item in selected_values:
                # Construct the one-hot encoded column name (e.g., "Spice_Level_Mild")
                one_hot_col_name = f"{df_col_prefix}_{val_item}"
                if one_hot_col_name in user_vector.index:
                    user_vector[one_hot_col_name] = 1.0

    # 3. Numerical Preferences in User Vector
    raw_user_num_prefs_dict = {}
    spice_map_for_num_pref = {'Mild': 1.0, 'Medium': 2.0, 'Hot': 3.0} # Use floats

    if 'price_range' in preferences and preferences['price_range']:
        try: # Average of min/max price for preference
            raw_user_num_prefs_dict['Price'] = (float(preferences['price_range'][0]) + float(preferences['price_range'][1])) / 2
        except: pass # Ignore if parsing fails
    if 'slices' in preferences and preferences['slices'] is not None:
        try: raw_user_num_prefs_dict['Slices'] = float(preferences['slices'])
        except: pass
    if 'rating' in preferences and preferences['rating'] is not None:
        try: raw_user_num_prefs_dict['Rating'] = float(preferences['rating'])
        except: pass
    if 'prep_time' in preferences and preferences['prep_time'] is not None:
        try: raw_user_num_prefs_dict['Preparation_Time'] = float(str(preferences['prep_time']).lower().replace("min","").strip())
        except: pass
    # Numerical Spice_Level: Only if *one* spice level is selected, use its mapped value.
    # Otherwise, rely on the one-hot encoded spice level features.
    if 'spicelevel' in preferences and isinstance(preferences['spicelevel'], list) and len(preferences['spicelevel']) == 1:
        selected_spice = preferences['spicelevel'][0]
        if selected_spice in spice_map_for_num_pref:
            raw_user_num_prefs_dict['Spice_Level'] = spice_map_for_num_pref[selected_spice]

    # Scale these raw numerical preferences using the SCALER
    # Create a temporary DataFrame for scaling, ensuring all NUMERICAL_COLS are present
    temp_scaling_df = pd.DataFrame(columns=NUMERICAL_COLS, index=[0])
    for col in NUMERICAL_COLS:
        # Default to the column's mean from FEATURE_DF if user didn't specify,
        # or 0 if that's also not available (shouldn't happen if SCALER is fit)
        # SCALER.data_min_ / SCALER.data_max_ or SCALER.mean_ could be used if available
        default_val = 0.0
        if hasattr(SCALER, 'data_min_') and col in FEATURE_DF.columns: # Check if scaler is fit and col exists
             # Use the minimum of the scaled range as a neutral default if user didn't specify
             col_idx_in_scaler = -1
             try: col_idx_in_scaler = NUMERICAL_COLS.index(col)
             except ValueError: pass

             if col_idx_in_scaler != -1 and col_idx_in_scaler < len(SCALER.data_min_):
                 default_val = SCALER.data_min_[col_idx_in_scaler] # This is the original min, not scaled min (0)
             else: # Fallback if col not in NUMERICAL_COLS used for SCALER fitting
                 logger.warning(f"Column {col} not found in SCALER's fitted columns during user vector creation. Defaulting to 0.")

        temp_scaling_df.loc[0, col] = raw_user_num_prefs_dict.get(col, default_val)


    if hasattr(SCALER, 'n_features_in_') : # Check if scaler has been fit
        scaled_user_num_values = SCALER.transform(temp_scaling_df[NUMERICAL_COLS])[0]
        for i, col_name in enumerate(NUMERICAL_COLS):
            if col_name in raw_user_num_prefs_dict: # Only update user_vector if user specified this preference
                user_vector[col_name] = scaled_user_num_values[i]
    else:
        logger.warning("SCALER is not fit. Cannot scale user's numerical preferences. Using raw values (0-1 range assumed).")
        for col_name in NUMERICAL_COLS:
            if col_name in raw_user_num_prefs_dict:
                 # Attempt a rough normalization if scaler is not fit, assuming values are in a reasonable range
                 # This is a fallback and might not be accurate.
                 user_vector[col_name] = raw_user_num_prefs_dict[col_name] / 100.0 # Example, needs domain knowledge


    # Calculate Cosine Similarities
    feature_matrix_filtered = filtered_feature_df.values
    user_array = user_vector.values.reshape(1, -1)

    # Ensure shapes match if FEATURE_DF columns changed dynamically (should not happen with current setup)
    if user_array.shape[1] != feature_matrix_filtered.shape[1]:
        current_app.logger.error(
            f"Shape mismatch! User vector: {user_array.shape}, Feature matrix: {feature_matrix_filtered.shape}. "
            f"User cols: {user_vector.index.tolist()[:5]}, Feature cols: {filtered_feature_df.columns.tolist()[:5]}"
        )
        # Attempt to align columns as a robust measure, though this indicates a deeper issue if it occurs.
        common_cols = filtered_feature_df.columns.intersection(user_vector.index)
        aligned_user_vector = pd.Series(0.0, index=filtered_feature_df.columns)
        aligned_user_vector[common_cols] = user_vector[common_cols]
        user_array = aligned_user_vector.values.reshape(1, -1)
        
        if user_array.shape[1] != feature_matrix_filtered.shape[1]:
            current_app.logger.critical(f"Persistent shape mismatch even after alignment. Cannot compute similarity.")
            return []


    similarities = cosine_similarity(user_array, feature_matrix_filtered)[0]
    # Get indices sorted by similarity (descending) from the filtered_feature_df
    sorted_indices_in_filtered_df = similarities.argsort()[::-1]
    # Map these sorted indices back to original DF indices
    final_recommendation_indices = valid_indices_for_feature_df[sorted_indices_in_filtered_df]

    # Prepare list of recommendations
    recommendations_list = []
    # frontend_keys and df_to_frontend_map are defined in index_route, can be reused or redefined here
    # For safety, redefine here or pass as argument if refactoring
    frontend_keys_rec = [
        'id', 'name', 'toppings', 'price', 'slices', 'serving_size', 'rating', 'rating_count',
        'description', 'popular_group', 'dietary_category', 'spice_level', 'sauce_type',
        'cheese_amount', 'calories', 'allergens', 'prep_time', 'restaurant', 'seasonal',
        'bread_type', 'image_url', 'crust_type'
    ]
    df_to_frontend_map_rec = {
        'id': None, 'name': 'Pizza_Name', 'toppings': 'Toppings', 'price': 'Price_Rs', 'slices': 'Slices',
        'serving_size': 'Serving_Size', 'rating': 'Rating', 'rating_count': 'Rating_Count',
        'description': 'Description', 'popular_group': 'Popular_Group',
        'dietary_category': 'Dietary_Category', 'spice_level': 'Spice_Level',
        'sauce_type': 'Sauce_Type', 'cheese_amount': 'Cheese_Amount',
        'calories': 'Calories_per_Slice', 'allergens': 'Allergens',
        'prep_time': 'Preparation_Time_min', 'restaurant': 'Restaurant_Chain',
        'seasonal': 'Seasonal_Availability', 'bread_type': 'Bread_Type',
        'image_url': 'Image_Url', 'crust_type': CRUST_TYPE_COL
    }

    for original_idx in final_recommendation_indices:
        pizza_series = DF.iloc[original_idx]
        rec_item = {}
        for key in frontend_keys_rec:
            df_col = df_to_frontend_map_rec.get(key)
            if key == 'id':
                rec_item[key] = int(original_idx)
            elif df_col and df_col in pizza_series:
                value = pizza_series[df_col]
                if isinstance(value, np.integer): value = int(value)
                elif isinstance(value, np.floating): value = float(value)
                elif isinstance(value, np.ndarray): value = value.tolist()
                rec_item[key] = "" if pd.isna(value) else value
            elif key == 'crust_type' and not CRUST_TYPE_COL :
                 rec_item[key] = "N/A"
            else:
                rec_item[key] = ""

        rec_item['rating_count'] = int(rec_item.get('rating_count', 0) or 0)
        rec_item['image_url'] = rec_item.get('image_url') if rec_item.get('image_url') else DEFAULT_IMAGE_URL
        for k_final, v_final in rec_item.items(): # Final numpy type check
            if isinstance(v_final, np.generic): rec_item[k_final] = v_final.item()
        recommendations_list.append(rec_item)

    current_app.logger.info(f"Final recommendations count: {len(recommendations_list)}")
    return recommendations_list


@app.route('/recommend', methods=['POST'])
def recommend():
    try:
        data = request.json
        preferences = {} # Store processed preferences
        current_app.logger.info(f"Received recommendation request with data: {data}")

        # Numerical/Range preferences from JS
        # Keys in `data` should match JS: 'slices', 'rating', 'prep_time', 'price_range'
        simple_numerical_prefs_js = ['slices', 'rating', 'prep_time']
        for key_js in simple_numerical_prefs_js:
            if key_js in data and data[key_js] is not None:
                try:
                    if key_js == 'rating': preferences[key_js] = float(data[key_js])
                    else: preferences[key_js] = int(data[key_js]) # slices, prep_time
                except ValueError:
                    current_app.logger.warning(f"Could not parse numerical preference '{key_js}': {data[key_js]}")
        
        if 'price_range' in data and data['price_range']:
            try:
                preferences['price_range'] = [float(p) for p in data['price_range']]
            except (ValueError, TypeError):
                 current_app.logger.warning(f"Could not parse price_range: {data['price_range']}")

        # Multi-select categorical preferences from JS
        # Keys in `data` should match JS: 'toppings', 'servingsize', 'dietarycategory', etc.
        multi_select_prefs_js = [
            'toppings', 'servingsize', 'populargroup', 'dietarycategory',
            'spicelevel', 'saucetype', 'cheeseamount', 'restaurantchain',
            'seasonalavailability', 'breadtype', 'crusttype'
        ]
        for key_js in multi_select_prefs_js:
            if key_js in data and isinstance(data[key_js], list):
                preferences[key_js] = data[key_js] # Expecting a list (can be empty for "Any")
            elif key_js in data: # If not a list, log warning
                current_app.logger.warning(f"Preference for '{key_js}' was not a list: {data[key_js]}. Treating as empty (Any).")
                preferences[key_js] = [] # Default to empty list if not a list

        current_app.logger.info(f"Processed preferences for filtering: {preferences}")
        recommendations = get_recommendations(preferences)
        current_app.logger.info(f"Returning {len(recommendations)} recommendations after filtering and scoring.")
        return jsonify(recommendations)

    except Exception as e:
        current_app.logger.error(f"Error in /recommend endpoint: {e}", exc_info=True)
        return jsonify({"error": "Failed to get recommendations due to a server issue.", "details": str(e)}), 500


# --- Main Application Execution ---
# Call preprocess_data() at the module level.
# This ensures it runs once when the application (or each Gunicorn worker) starts.
try:
    logger.info("----- Starting data preprocessing at module load... -----")
    preprocess_data() # Use default 'pizza.csv'
    logger.info("----- Data preprocessing completed successfully at module load. -----")
    if DF is None:
        logger.critical("CRITICAL AT STARTUP: Global DF is None after preprocess_data(). App will likely fail.")
    if FEATURE_DF is None:
        logger.critical("CRITICAL AT STARTUP: Global FEATURE_DF is None after preprocess_data(). App will likely fail.")
    if SCALER is None: # SCALER should be initialized even if fitting fails
        logger.critical("CRITICAL AT STARTUP: Global SCALER is None after preprocess_data(). App will likely fail.")

except FileNotFoundError as e:
    logger.critical(f"CRITICAL ERROR AT MODULE LOAD (FileNotFoundError): {e}. Ensure 'pizza.csv' is in the /app directory (or same dir as app.py).")
    # In a production Gunicorn setup, the app might still try to start, leading to errors in routes.
    # For Hugging Face, it's better to log and let it attempt to run, as exiting might obscure logs.
except Exception as e:
    logger.critical(f"Unexpected critical startup error during preprocessing at module load: {e}", exc_info=True)


if __name__ == '__main__':
    # This block is primarily for local development using `python app.py`.
    # preprocess_data() is already called above when the module is imported by Python interpreter.
    logger.info("----- Running Flask app directly (e.g., python app.py) -----")
    # Sanity check for local run, though globals should be set by the module-level call.
    if DF is None or FEATURE_DF is None or SCALER is None:
        logger.warning("One or more global data variables (DF, FEATURE_DF, SCALER) are None before local app.run(). This is unexpected if module-level preprocessing ran.")
        # Optionally, re-run preprocessing if critical for local dev and something went wrong with module-level load
        # logger.info("Attempting to re-run preprocess_data() for local development.")
        # preprocess_data()

    app.run(debug=True, host='0.0.0.0', port=7860, use_reloader=False)
    # use_reloader=False is generally better when you have global state initialized at module level.
    # If True, it might re-initialize globals on each reload, which can be slow.