ocr-and-documents

Use pymupdf to open the PDF and call get_text() on each page, or pymupdf4llm for markdown output. The extract_pymupdf.py script handles text, markdown, tables, images, and metadata via command-line flags.

pymupdf is lightweight (~25MB) and instant, handling text-based PDFs, tables, and images. marker-pdf (~3-5GB with PyTorch) adds OCR for scanned documents, equations, forms, and reading-order detection. Use pymupdf unless you need OCR or complex layout analysis.

Scanned PDFs require OCR since they contain images rather than text layers. Use marker-pdf, which supports OCR in 90+ languages, via the extract_marker.py script. pymupdf cannot perform OCR on scanned documents.

Yes, marker-pdf needs approximately 5GB free for PyTorch plus models, and downloads ~2.5GB of models to ~/.cache/huggingface/ on first use. Run extract_marker.py --check to verify disk space before installing.

Yes, use web_extract with the document URL first, which converts PDFs to markdown via Firecrawl with no local dependencies. Only fall back to local extraction when the file is local, web_extract fails, or batch processing is needed.

pymupdf handles splitting and merging natively with no extra dependencies. Use insert_pdf to copy page ranges into a new document for splitting, or combine multiple source documents into one for merging, then save the result.