Extracting with a vision model

You've now seen modular tools (PyMuPDF + Tesseract) and combined packages (Docling). Both work at the character level — reading shapes and matching patterns. Vision-capable LLMsA model trained on text to predict the next token, and so to generate language. take a fundamentally different approach: they interpret page images directly, understanding layout, context, abbreviations, and even handwriting.

This is the most capable extraction method available — and also the most expensive and slowest.