OCR (optical character recognition)
OCR — optical character recognition — converts images of text, such as scanned pages or photographs, into machine-readable characters that can be searched, selected and quoted.
- Also called
- Text recognition
- Dangerous case
- A page that is 70% right
- Report quality
- Per page, not per document
The failure mode is not the one people expect
Running OCR is a solved-enough problem. The part that damages your work is quality assessment, because bad recognition rarely looks bad.
A page that fails completely is harmless: you see nonsense, you know not to trust it. A page that comes out 70% right is the dangerous one. It reads fine. You quote from it. Three weeks later someone points out that the quoted sentence does not appear in the book, because a digit flipped, a negation dropped, or two columns interleaved into a sentence that never existed.
This is why a single document-level accuracy figure is close to useless. “94% accurate” tells you nothing about which 6%, and the 6% is entirely what you needed to know.
What actually goes wrong, in rough order of frequency:
- Layout, not characters. Multi-column pages and tables often produce correct letters in wrong order. The text looks plausible and means something different.
- Physical scan quality. Skew, tight gutters on bound books, bleed-through from the reverse side, low-resolution photocopies. None of this is recoverable after the fact — the information was never captured.
- Handwriting and marginalia. Unreliable enough to treat as a hint rather than as text.
- Notation. Mathematics, chemical formulae and non-Latin scripts mixed into Latin text all degrade sharply.
How MindMax handles it
Two decisions, both about honesty rather than accuracy.
Recognition is something you start, not something that happens. It runs on the document you chose, not silently across your folder — because it costs real compute and you are the one who knows which book is worth it.
Quality is reported per page, not per document. Each page gets its own record and is flagged when the result looks wrong. The check is on the shape of the output rather than on finding expected words, so it behaves the same on a language it has never seen. You can then re-run recognition on just the flagged pages instead of the whole book.
- Search it
- Highlight it
- Ask about it
- Quotes land on the page
Detected as a scan, recognised when you say so, checked page by page — and the recognised text points back at the page image it came from.
The point is not that MindMax recognises text better than other tools. It is that you are told where not to trust it — which is the information that decides whether a scanned source is safe to cite.
After a run: how many pages were recognised and how many need review, with the per-page report one click away.
Common questions
How accurate is OCR?
On a clean, single-column printed page, very. Accuracy drops with multi-column layouts, tables, poor scans, handwriting and mathematical notation — and a document-wide percentage hides which pages were affected. A per-page quality record is what tells you which passages to double-check.
How can I tell whether OCR got a page wrong?
Look for structural signs rather than spelling: lines in the wrong order, columns run together, unusually short or fragmented output. MindMax checks each page’s output for these and flags the pages that look wrong, so you can re-run or check just those.
Can I ask an AI about a scanned PDF?
Once its text has been recognised, yes — the recognised text is used for questions, annotations, search and translation, and quotes point back at the page. Before recognition, MindMax tells you the document is a scan rather than answering from images it cannot read as text.