Searching inside a scanned document, with no connection at all

What text recognition actually does

When you take a photo of a printed or handwritten page, the file itself is still just an image: a grid of pixels with no idea what any of it says. Text recognition, usually called optical character recognition or OCR, is the separate step that looks at those pixels and works out which of them form letters and words. Once that step has run, the words exist as real text alongside the image, and a search field can look for them the way it looks for a word in an email.

Apple provides on-device text recognition to apps through a framework called Vision. People meet a related, system-level version of it directly as Live Text: the thing that lets you tap and copy a phone number or an address straight out of a photo in the Photos or Camera app. Any app can use the Vision framework to run its own text recognition on the device, though that does not mean every app behaves identically — how thorough or fast the recognition is still depends on how the app uses it. Apple documents that this recognition runs on the device itself; what an app then does with the original image afterwards, including whether it uploads it anywhere, is up to that app. Pocket Aurora Papers does not upload page images in this version.

From recognised text to a search you can actually use

Recognising the words on a page is necessary, but it is not the same as being able to find them later. A search field that is fast needs an index: a separate structure that has already broken the recognised text into searchable pieces, so it doesn't have to re-read every page from scratch each time you type. Pocket Aurora Papers shows this step directly: a document’s Status reads Reading… for a moment after a scan, then Read once its text is ready and searchable, or No text found if a page genuinely had none. That gap is the index being built, not the recognition itself, and it is worth expecting rather than treating as a fault.

Why running on the device, not a server, matters

Two things follow from recognition happening on the phone itself rather than after an upload. The first is speed: there is no waiting on a network round trip before you find out whether the page came through legibly. The second, and the one that matters most for paperwork, is that it keeps working when there is no connection at all: on a train, in a basement flat with no signal, or with the phone in airplane mode. A useful practical check, though not formal proof, is the airplane-mode test: if scanning and searching still work with connectivity switched off, that is good evidence the recognition step itself does not depend on a server somewhere to do the reading for you.

What makes a scan easy or hard to read

Recognition quality is not a fixed constant: how well it works on any given page depends on the image it is given — its lighting, its layout, whether the page is flat or curled — and on whether the text is written in a language the recognition step actually supports. A handful of habits make a real difference to how much of a page comes back as usable text.

• Light the page evenly. A single lamp to one side throws a shadow across half the text; daylight from a window, or two light sources, spreads more evenly.
• Keep the page flat. A curled or bent corner distorts the letters along that edge, and recognition struggles with warped text even when a human eye could still read it.
• Fill the frame with the whole page, not more and not less. Cropping into the text loses words at the edges; leaving too much table or background around it can confuse automatic edge detection.
• Hold steady for the moment of capture. A small amount of motion blur affects recognition far more than it affects how a photo looks to the eye at a glance.
• Avoid glare, especially on laminated cards or plastic-sleeved documents; angling the page slightly, rather than the camera, usually clears it.

None of this needs special equipment. It is closer to the difference between a photo you would send a friend and a photo taken with the specific job of being read by a machine afterwards.

How Pocket Aurora Papers handles this

Pocket Aurora Papers reads text on the device itself, in English, Czech, German, French and Spanish, and then makes that text searchable together with a document's title and tags. Because recognition and the search index both run locally, scanning, reading and searching your archive all work with no connection at all. When a page comes out blurred or badly lit, the app tells you rather than guessing: each page is checked as you take it, as a good scan, one worth a second look, or one worth retaking, and none of them is ever blocked. You choose whether to retake it, adjust its corners or keep it as it is, since only you know whether the page was good enough for what you need it for. See how the archive works for the full sequence from capture to search, and what we store, and what we don't for exactly what stays on the device and what only happens if you sign in.

Search itself also has to cope with accents typed the fast way and with languages like German that build long compound words. That is its own subject — see why document search should ignore accents and find compound words.