perf: load seed data from Parquet instead of generating on every query

app.py

@@ -14,10 +14,24 @@ from model_inference import load_model, generate_sql
 from data_engine import create_session, execute_safe, QueryTimeoutError
 
 # ── Startup ───────────────────────────────────────────────────────────
-# Model loads into RAM at startup; Zero GPU Spaces run CPU-only so the
-# model stays resident between @spaces.GPU calls.
 
 print("🚀 Starting Kasualdad LFED...")
+
+# Ensure Parquet seed files exist (generate on first boot, persist in /data/)
+from pathlib import Path
+_parquet_dirs = [Path("/data"), Path(__file__).parent / "data"]
+_pq_files = ["enrollment.parquet", "attendance.parquet"]
+_pq_found = any(
+    all((base / f).exists() for f in _pq_files)
+    for base in _parquet_dirs
+)
+if not _pq_found:
+    print("📦 Generating seed Parquet files (first boot)...")
+    from data.export_parquet import export_parquet
+    _pq_out = _parquet_dirs[0] if _parquet_dirs[0].exists() else _parquet_dirs[1]
+    export_parquet(_pq_out)
+
+print("🦙 Loading model...")
 llm = load_model()
 print("✅ Ready.")
 

data/export_parquet.py

@@ -0,0 +1,65 @@
+"""
+export_parquet.py — Generate seed Parquet files for Kasualdad LFED.
+
+Run once to produce enrollment.parquet and attendance.parquet.
+Place them in /data/ (HF Space persistent storage) for fast loading.
+
+Usage:
+    python data/export_parquet.py                    # → data/
+    python data/export_parquet.py --out /data        # → /data/ (HF Space)
+"""
+
+import argparse
+import sys
+from pathlib import Path
+
+# Allow running from repo root
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+import duckdb
+from data.generate_seed import generate_seed_data
+
+
+def export_parquet(out_dir: Path) -> None:
+    """Generate seed data and export both tables as Parquet."""
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    conn = duckdb.connect(":memory:")
+    conn.execute("SET enable_progress_bar = false;")
+    generate_seed_data(conn)
+
+    enrollment_path = out_dir / "enrollment.parquet"
+    attendance_path = out_dir / "attendance.parquet"
+
+    conn.execute(f"COPY enrollment TO '{enrollment_path}' (FORMAT PARQUET)")
+    conn.execute(f"COPY attendance TO '{attendance_path}' (FORMAT PARQUET)")
+
+    size_enroll = enrollment_path.stat().st_size
+    size_attend = attendance_path.stat().st_size
+
+    print(f"✅ Exported:")
+    print(f"   {enrollment_path}  ({size_enroll:,} bytes)")
+    print(f"   {attendance_path}  ({size_attend:,} bytes)")
+    print(f"   Total: {(size_enroll + size_attend):,} bytes")
+
+    # Verify round-trip
+    verify = duckdb.connect(":memory:")
+    verify.execute(f"CREATE TABLE enroll AS SELECT * FROM read_parquet('{enrollment_path}')")
+    verify.execute(f"CREATE TABLE attend AS SELECT * FROM read_parquet('{attendance_path}')")
+    n_enroll = verify.execute("SELECT COUNT(*) FROM enroll").fetchone()[0]
+    n_attend = verify.execute("SELECT COUNT(*) FROM attend").fetchone()[0]
+    print(f"   ✓ Round-trip OK: {n_enroll} enrollment rows, {n_attend} attendance rows")
+    verify.close()
+    conn.close()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Export seed data to Parquet")
+    parser.add_argument(
+        "--out",
+        type=Path,
+        default=Path(__file__).parent,
+        help="Output directory (default: data/)",
+    )
+    args = parser.parse_args()
+    export_parquet(args.out)

data_engine.py

@@ -34,6 +34,15 @@ QUERY_TIMEOUT_SEC = 10
 DATA_DIR = Path(__file__).parent / "data"
 SEED_SQL_PATH = DATA_DIR / "seed.sql"
 
+# Parquet files (pre-generated once via data/export_parquet.py).
+# On HF Spaces, place them in the persistent /data/ directory.
+_PARQUET_DIRS = [
+    Path("/data"),          # HF Space persistent storage
+    DATA_DIR,               # local dev (data/)
+]
+_ENROLLMENT_PQ = "enrollment.parquet"
+_ATTENDANCE_PQ = "attendance.parquet"
+
 
 # ── Connection factory ─────────────────────────────────────────────────
 
@@ -52,28 +61,50 @@ def get_connection(read_only: bool = False) -> duckdb.DuckDBPyConnection:
 
 # ── Database seeding ───────────────────────────────────────────────────
 
-def seed_database(conn: duckdb.DuckDBPyConnection, seed_sql_path: Path | None = None) -> None:
+def seed_database(
+    conn: duckdb.DuckDBPyConnection,
+    seed_sql_path: Path | None = None,
+) -> None:
     """
-    Create tables and load seed data into the given connection.
+    Create tables and load seed data.  Tries, in order:
 
-    Tries data/seed.sql first (for custom overrides), otherwise uses the
-    Python seed data generator (5 schools × 4 years, ~500 students).
+      1. Parquet files (fastest — pre-generated, ~33 KB total)
+         → /data/enrollment.parquet + /data/attendance.parquet (Space)
+         → data/enrollment.parquet + data/attendance.parquet (local)
+      2. data/seed.sql (custom overrides)
+      3. Python generator (slow fallback)
     """
+    # ── 1. Parquet files ─────────────────────────────────────────
+    for base in _PARQUET_DIRS:
+        enroll_pq = base / _ENROLLMENT_PQ
+        attend_pq = base / _ATTENDANCE_PQ
+        if enroll_pq.exists() and attend_pq.exists():
+            conn.execute(
+                "CREATE TABLE enrollment AS "
+                f"SELECT * FROM read_parquet('{enroll_pq}')"
+            )
+            conn.execute(
+                "CREATE TABLE attendance AS "
+                f"SELECT * FROM read_parquet('{attend_pq}')"
+            )
+            return
+
+    # ── 2. seed.sql ──────────────────────────────────────────────
     if seed_sql_path is None:
         seed_sql_path = SEED_SQL_PATH
 
     if seed_sql_path.exists():
-        # User-provided seed.sql — execute it
         with open(seed_sql_path) as f:
             sql = f.read()
         for statement in sql.split(";"):
             statement = statement.strip()
             if statement and not statement.startswith("--"):
                 conn.execute(statement)
-    else:
-        # Default: use Python generator with realistic distributions
-        from data.generate_seed import generate_seed_data
-        generate_seed_data(conn)
+        return
+
+    # ── 3. Python generator (slow) ───────────────────────────────
+    from data.generate_seed import generate_seed_data
+    generate_seed_data(conn)
 
 
 # ── Schema introspection ───────────────────────────────────────────────