Making a scanned document searchable
A scanned PDF is a picture. You can read it, but your computer cannot: search finds nothing, text cannot be copied, and a screen reader has nothing to read. For an archive of scanned records this is a serious limitation — the information is present and effectively unfindable.
Optical character recognition examines the image, identifies the shapes as characters, and adds a text layer beneath the picture. The document looks exactly the same and becomes searchable, selectable and accessible.
The practical value grows with volume. One scanned letter is easy to read; two hundred scanned letters without OCR are a filing cabinet you have to go through by hand, and with OCR they are a searchable archive.
What determines accuracy
Scan quality dominates everything else. A clean, straight scan of printed text at 300 DPI recognises with very high accuracy; the same document photographed at an angle in poor light does not. Skew is particularly damaging, because character recognition assumes roughly horizontal text lines.
Typefaces matter too. Standard printed fonts recognise well. Decorative fonts, dot-matrix output, faxes that have been through several generations, and handwriting range from unreliable to hopeless — handwriting recognition is a genuinely different and much harder problem than printed-text OCR.
Layout complexity is the third factor. Single-column text is straightforward. Multi-column pages, tables, sidebars and text wrapped around images all require the software to work out reading order, and mistakes there produce text that is individually accurate and collectively scrambled.
Checking the result before relying on it
Never assume OCR output is correct, particularly for numbers. The characters most often confused are exactly the ones where an error matters: zero and the letter O, one and the letter l, five and S, eight and B. In an address or a reference number those substitutions are easy to miss and consequential.
Search the finished document for a few terms you know are in it. If they are found, the text layer is working; if a common word returns nothing, something has gone wrong and it is better to know now.
One privacy point worth stating plainly, since it is the reason this runs locally. OCR requires reading every word of the document, and the documents people OCR are typically archives of personal or business records. Sending those to a cloud service for processing is a substantial disclosure — here the recognition engine runs in your browser, so the contents stay on your machine.
Getting the best recognition from what you have
If you control the scan, a few settings dominate the outcome. Scan at 300 DPI — higher rarely improves recognition and slows everything down, while lower loses the detail that distinguishes similar characters. Scan in greyscale rather than colour for text documents, and avoid the aggressive black-and-white modes some scanners default to, which tend to break thin strokes.
Straighten before recognising. Even a few degrees of skew measurably reduces accuracy, because the software assumes text lines run horizontally. Most scanning software will deskew automatically, and doing it before OCR rather than after is what matters.
If you only have a photograph, use your phone's document mode rather than the plain camera. It corrects perspective and evens out lighting, which addresses the two things that most damage recognition. Failing that, retaking the photograph in better light is almost always faster than correcting the text afterwards.
Set the recognition language to match the document rather than leaving it at the default. Language selection influences how ambiguous shapes are resolved, because the engine weighs candidates against the words and character patterns that language actually uses. A document in a language the engine was not told to expect recognises measurably worse, and the errors it produces tend to be plausible words rather than obvious nonsense, which makes them harder to spot on review.