text-layer (PyMuPDF) → OCR fallback (Tesseract ind+eng) → LLM parsing (structured JSON) + cache learning.