ocr-and-documents

Use pymupdf to open the PDF and call get_text() on each page, or pymupdf4llm to convert directly to Markdown. It installs in about 25MB and handles text-based PDFs, tables, images, and metadata without downloading models.

pymupdf is lightweight and instant, best for text-based PDFs, basic tables, and image extraction. marker-pdf handles scanned documents with OCR in 90+ languages, equations, forms, and reading order detection, but requires 3-5GB for PyTorch and models.

Use marker-pdf, which performs OCR on scanned PDFs across 90+ languages and outputs Markdown with layout analysis. Run the extract_marker.py script on the file, optionally with --use_llm for higher accuracy or --json for structured output.

No, pymupdf cannot OCR scanned PDFs or convert images to text; it only extracts existing text layers and embedded images. For scanned documents, use marker-pdf or provide a URL so web_extract can handle the conversion.

marker-pdf needs roughly 5GB of free disk for PyTorch plus about 2.5GB of models downloaded to the HuggingFace cache on first use. Processing takes 1-14 seconds per page on CPU, so check disk space with the --check flag before installing.

pymupdf handles split and merge natively with no extra dependencies. Open the source PDF, create a new document, use insert_pdf with page ranges to split, or insert multiple documents sequentially to merge, then save the result.