ocr-and-documents

Use pymupdf to open the PDF and call get_text() on each page, or pymupdf4llm for Markdown output. For remote PDF URLs, try web_extract first since it handles PDF-to-markdown conversion with no local dependencies.

pymupdf is lightweight at ~25MB, runs instantly, and handles text-based PDFs, tables, and images. marker-pdf needs ~3-5GB for PyTorch and models but adds OCR for scanned documents, equations, forms, and reading order detection.

Scanned PDFs require OCR, which pymupdf does not support. Use marker-pdf, which performs OCR in over 90 languages and outputs Markdown with layout analysis, at roughly 1-14 seconds per page on CPU.

Open the source PDFs with pymupdf, create a new empty document, and use insert_pdf() to copy selected pages or entire files, then save the result. No extra dependencies are needed beyond pymupdf itself.

marker-pdf requires ~5GB of free disk for PyTorch plus ~2.5GB of models downloaded to the HuggingFace cache on first use. Run the extract_marker.py --check flag first to verify disk space, or fall back to pymupdf for text-based PDFs.