Document OCR (docling + vision-language model)

Purpose

Turn scanned PDFs and document images into clean, structured Markdown with figures preserved. The pipeline combines:

1. docling — detects layout (figures, tables, multi-column reading order) on each page, entirely on CPU.
2. A vision-language model (VLM) — transcribes each full page to Markdown, preserving reading order, diacritics, ligatures, special characters (♂ ♀ ½ æ, Greek/Latin), equations (LaTeX), and tables (HTML).

docling supplies reliable figure/table crops; the VLM supplies high-quality text. Born-digital PDFs are detected and extracted directly (no server needed).

Adapted from Bruno de Medeiros' OntoMorphoGrapher proof-of-concept, where the olmocr-docling backend processed an 18-PDF historical-entomology corpus (1833–2015, 5 languages) with no catastrophic failures.