Question 1

How is this different from every other OCR tool?

Accepted Answer

Where the work happens. Recognition is the one PDF feature the whole industry runs on its own servers, so making a scan searchable means sending it somewhere — and the scans people need this for are contracts, invoices, medical records and certificates. Here the model runs in your browser: it is downloaded once, about 21 MB, and after that the file never leaves the machine. You can watch the network panel while it works and see nothing being sent.

Question 2

How accurate is it, honestly?

Accepted Answer

Good, and not guaranteed. On a sixteen-page test set rendered with a 0.3 degree skew — because a real scan always has some — character recall measured 95.2%. Clean, straight pages score higher; stamps, handwriting, faint photocopies and pages photographed at an angle score lower, sometimes much lower. Treat it as making a document findable rather than as a transcript: the words you search for will usually be there, and a legal-grade transcription this is not.

Question 3

Does the scan itself change?

Accepted Answer

No, and this is checked rather than asserted: after a run, the SHA-256 of every image stream in the output matches the input exactly. The pages are never decoded, re-encoded or resized — only a text layer and its font description are added, which came to 5.1 KB on a two-page scan. If you also want the file smaller, compress it afterwards rather than instead: compression re-encodes the images, and doing that first would give the recogniser less to read.

Question 4

Can I see or select the added text?

Accepted Answer

You cannot see it, and you can select it. The text is written in the rendering mode that paints nothing, positioned over the words it came from, so the page looks untouched while search, copy and screen readers all find the words. That also means no font is embedded — the glyphs are never drawn, so only the character mapping is needed, and the font costs zero bytes. The output was read back through two independent PDF engines rather than only the one we use, and both returned the same Chinese text.

Question 5

What if the PDF already has text in it?

Accepted Answer

It is left alone and the tool says so. A PDF exported from Word is already searchable, and putting a recognition layer on top would make every search match twice and every copy come out doubled. What the tool checks is whether the pages carry text at all, sampled through the document; if they do, there is nothing here to add.

Question 6

How long does it take?

Accepted Answer

A few seconds per page for an ordinary Chinese or English page — measured at 2.4 to 4.0 seconds on a desktop with hardware acceleration — plus the one-time model download. A dense two-column academic page, four thousand characters, took about fourteen. The cost tracks how much text is on the page rather than the page size, and a phone will be several times slower.

OCR PDF

FAQ