What is OCR?
The step that turns pictures of pages back into text. Decades old, still the gate between archives and search.
OCR, optical character recognition, reads text out of images of pages. A scanned PDF is a photograph, and a tool that searches photographs finds nothing; OCR converts the picture back into text so it can be searched, quoted and cited like any born-digital document. The software is decades old and runs comfortably on your own machine, which means scans of confidential material never need an upload to become searchable.
Pictures of pages are not text.
A born-digital PDF carries its text inside it, and a tool can lift that text out without doing anything clever. A scanned PDF carries only an image, pixels arranged in the shape of letters. Until something reads those pixels, the document is invisible to every search box, which is why archives of court filings, old journal articles and scanned theses sit unreachable in folders. OCR closes that gap: it looks at the image and produces the text the marks represent, stored alongside the page as a searchable layer. That text is what a research tool indexes, which is how scans join the same research loop as everything else in a library.
Where it fails.
OCR is very good at exactly one thing: clean, printed, upright text. It degrades on tilted scans, faint photocopies, tight columns, tables, and handwriting, and it can fail silently, returning confident text with wrong characters rather than reporting that it could not read the page. Vision models take a different route, reading the page image directly, and handle layouts that trip plain OCR. Neither excuses the reader from checking: extracted text that you intend to quote deserves a look against the page image, the same way any citation does. The tool makes the text reachable; the judgment stays with you.
Why it can stay local.
OCR is not a frontier-model problem. The engines are small, well-understood software, and modern operating systems ship one: Windows includes an OCR engine that a local tool can call with no download, no account and no upload. The usual cloud pattern, sending scans to a service that reads them, is a habit of convenience, not a requirement. For anything confidential, from unpublished clinical work to a thesis before submission, keeping the conversion on your own machine is the difference between a searchable archive and a leak. It is one more part of the loop that never has to leave the disk.