OCR and scanned documents

Use this Skill when ordinary document extraction fails because the source is scanned, image-based, or layout-heavy.

Start here

Call skill_view for this Skill and use its returned baseDirectory. Helpers live in /scripts.

Prefer local processing. Do not upload a private document to an external OCR service unless the user explicitly authorizes that transfer.

Choose the lightest path

1. Normal extraction first. For PDFs, try the bundled pdf Skill. For .docx, use docx. Do not OCR born-digital text.
2. A few scanned pages. Render pages with the pdf Skill and inspect images using view_image; transcribe only the requested pages.
3. Text-oriented scanned PDF. Try extract_pymupdf.py; this is the lighter local path.
4. Complex layout, tables, math, or many scanned pages. Use extract_marker.py only after confirming the need and checking resources.

Helpers

python3 <baseDirectory>/scripts/extract_pymupdf.py input.pdf -o extracted.md
python3 <baseDirectory>/scripts/extract_marker.py --check
python3 <baseDirectory>/scripts/extract_marker.py input.pdf -o extracted.md

Run helpers with terminal_exec, using python3 as the command and arguments as an array. Use --help for exact options.

extract_pymupdf.py requires PyMuPDF:

python3 -m pip install pymupdf

Its --markdown mode additionally requires pymupdf4llm.

extract_marker.py uses marker-pdf, which can require several gigabytes of packages and model downloads. Never install it automatically. Check available disk space and ask the user before installing or downloading large models.

URLs

If the source is a URL, first use an available browser or HTTP tool to identify whether it points to HTML or a document. Download only the requested document, respect authentication and access boundaries, and then process the local copy. Do not assume a specific web tool exists.

Verification

• Preserve reading order, page references, headings, tables, footnotes, and confidence caveats.
• Compare extracted text against representative rendered pages with view_image.
• Mark uncertain characters and fields instead of guessing.
• For tables, verify column alignment and numeric values separately.
• Return both the extracted artifact path and the original source/page mapping when possible.