src/state.py

class AppState:
    df = None
    node_features = None
    graph = None
    louvain_partition = None
    pagerank_scores = None
    betweenness_scores = None
    full_features = None
    alerts = []
    xgb_bundle = None
    model_metrics = None
    gnn_metrics = None

    # Pre-built caches for fast API responses
    features_by_account = {}      # account_id -> feature dict (O(1) lookup)
    cached_channel_stats = []     # pre-computed channel breakdown
    cached_overview = {}          # pre-computed overview numbers
    cached_typo_counts = {}       # pre-computed alerts typologies
    cached_crit_alerts = 0        # pre-computed critical count
    accounts_by_number = {}       # account_number -> {bank_name, bank_id, entity_id, entity_name} (O(1) lookup from accounts.csv)

    # Entity/bank network intelligence — built once from accounts_by_number (no second CSV read)
    entities_by_id = {}           # entity_id -> {entity_id, entity_name, entity_type, account_numbers, bank_ids, bank_names}
    banks_by_id = {}              # bank_id -> {bank_id, bank_name, country_label, account_numbers, entity_ids}
    network_summary_cache = {}    # pre-computed /network/summary payload
    bank_profiles_cache = {}      # bank_id -> pre-computed /bank/{id}/profile payload
    clusters_by_id = {}           # community_id -> pre-computed /network/cluster/{id} aggregate (Louvain-derived "mule ring" groups)
    alerts_by_account = {}        # account_id -> [alert dict, ...] (O(1) index into AppState.alerts for entity-level typology breakdowns)

    # Uploaded-dataset tracking — when the user ingests their own data via the
    # Upload page, validation metrics (F1, AUC-ROC, confusion matrices) are no
    # longer meaningful because the upload carries no ground-truth labels.
    is_uploaded_dataset = False   # True once a user dataset replaces the IBM data
    uploaded_dataset_name = None  # original filename of the uploaded dataset
    uploaded_has_labels = False   # True if the upload included an 'Is Laundering' column

    # Startup tracking — lets the frontend show a loading screen
    startup_ready = False   # dashboard data ready
    graph_ready = False     # graph built; investigation/subgraph endpoints available
    startup_status = {
        'current_step': 'Initializing...',
        'steps_done': 0,
        'total_steps': 7,
        'errors': []
    }