ocr-and-documents

Use marker-pdf for scanned PDFs since they have no text layer for standard parsers to read. It performs OCR in over 90 languages and outputs Markdown, but requires roughly 3-5GB of disk space for PyTorch and models.

pymupdf is lightweight (~25MB) and instant, handling text-based PDFs, tables, and metadata. marker-pdf (~3-5GB) adds OCR, equation and LaTeX recognition, form handling, and reading-order detection for complex or scanned documents.

Use web_extract on the URL first, which converts remote PDFs to Markdown via Firecrawl with no local dependencies. Only fall back to local extraction with pymupdf or marker-pdf when the file is local or web_extract fails.

Empty output means those pages are scanned images without a text layer. For a few pages, render them with pdftoppm and analyze with a vision model; for bulk OCR across many pages, use marker-pdf.

Yes, pymupdf handles splitting, merging, and text search natively with no extra dependencies. Use insert_pdf to copy page ranges between documents and page.search_for to locate text across all pages.

pymupdf cannot perform OCR on scanned images, recognize equations or LaTeX, handle forms, or detect reading order. For those capabilities, switch to marker-pdf, which requires significantly more disk space and processing time.