OCRing PDFs

ocrmypdf writes an invisible text layer over the original page images, so the output file is both the scan you can look at and a document pdftotext, grep, and pdfplumber can read. Rasterize-then-tesseract gives you a .txt divorced from the pages; page numbers and coordinates are gone.

The toolchain is not in the base container. It installs in 18 seconds (measured 2026-09-12: apt 3s, pip 15s), so install it when a scan shows up rather than carrying it in a container layer.

Probe before installing

pdftotext in.pdf - | tr -d '\f \n' | wc -c

Nonzero means the PDF already has a text layer and is not a scan. Extract with pdftotext or pdfplumber and stop. Running OCR on it wastes a minute, and with --force-ocr it replaces exact embedded text with a lossy reading of a raster of itself.

A small nonzero count (tens of characters across many pages) is the mixed case: a born-digital cover page in front of scanned body pages, or a scan whose producer stamped a header. --skip-text handles it.

Install

sh scripts/ensure_ocr.sh              # English
sh scripts/ensure_ocr.sh nor deu      # plus Norwegian and German

Idempotent: 0.8s when everything is already present, 2.6s to add one more language pack. Installs ghostscript, pngquant, poppler-utils, tesseract and its language packs via apt, then ocrmypdf via pip.

Run

ocrmypdf --skip-text --deskew --rotate-pages --output-type pdf in.pdf out.pdf
pdftotext out.pdf - | wc -w        # verify: zero words means it failed quietly

About 2s per page for a single dense page at 200 DPI on one core. A 300-page scan is therefore a background job, not a single bash call — launch it detached with a sentinel file per the external-call pattern in bash-tool-timeout.

Which text-layer mode

flag	use it when
--skip-text	Default. Pages that already carry text are passed through untouched; image-only pages get OCR. The safe choice for anything mixed.
--force-ocr	Every page is rasterized and re-OCRed, discarding any existing text. Correct for a scan carrying a junk text layer, and for pages with text-over-image that --skip-text would skip. Destroys real embedded text, so probe first.
--redo-ocr	Replaces a previous OCR layer while leaving born-digital text alone. Narrower than --force-ocr and slower to fail on odd inputs.

--output-type pdf skips PDF/A conversion. Drop it when the output is going into an archive that requires PDF/A; ghostscript does the conversion either way.

Languages

-l eng+nor for a mixed-language document, -l nor for a monolingual one. Order does not matter. Every code needs its tesseract-ocr-<code> pack installed. Pass the codes to ensure_ocr.sh and it handles them. Accuracy drops noticeably when the language is wrong, and tesseract will not tell you; it returns confident garbage instead.

Container facts (measured 2026-09-12)

• apt-get update exits 100 here. A preconfigured nodesource repo is off the egress allowlist and returns 403, and the nonzero exit aborts any && chain behind it. The Ubuntu mirrors are reachable without an update. Run apt-get install directly.
• unpaper is absent, so --clean and --clean-final fail. Don't pass them.
• One core, so --jobs buys nothing on claude.ai. CCotw has four.
• ocrmypdf --version prints to stderr. Capture with 2>&1 or a version check reads as empty.
• jbig2 is absent; output uses CCITT/JPEG instead, which costs some file size and nothing else.

When to use transcribing-images instead

This skill produces glyphs. It does not read a chart, describe a diagram, or recover handwriting. Tesseract on those pages returns nothing useful and gives no sign that it lost anything.

Route to transcribing-images when the meaningful content is a picture, or when the deliverable is a reading rather than a file. Both is a normal answer: OCR the document so it is greppable, then send the pages that carry figures to a vision model.

In an interactive session, native vision beats both for a handful of pages: rasterize with pdftoppm -r 200 -png and view the images. Reach for OCR when the document is longer than context will hold, or when the text has to outlive the conversation as a file.