ocr-and-documents

Use pymupdf to open the PDF and call get_text() on each page, or run the extract_pymupdf.py script for plain text, Markdown, tables, or metadata. It installs in seconds at about 25MB and handles text-based PDFs instantly.

pymupdf is lightweight (~25MB) and instant but only reads existing text layers. marker-pdf (~3-5GB with PyTorch) adds OCR for scanned documents in 90+ languages, plus equations, forms, and reading-order detection. Use pymupdf unless you need OCR or complex layout analysis.

Use marker-pdf via the extract_marker.py script, which performs OCR on scanned pages and outputs Markdown. For just a few pages, render them with pdftoppm at 150 DPI and analyze each image with a vision model instead.

No, pymupdf only reads the existing text layer and returns nothing for scanned image pages. For those documents you need marker-pdf's OCR, or render pages to images and use a vision model for small page counts.

marker-pdf requires roughly 5GB of free disk for PyTorch plus about 2.5GB of models downloaded to the HuggingFace cache on first use. Run extract_marker.py --check to verify disk space, or fall back to pymupdf or web_extract for URL-based documents.

Call web_extract with the PDF URL first, which converts it to Markdown via Firecrawl with no local dependencies. This works for arXiv papers and hosted reports; only use local extraction when the file is local or web_extract fails.