OCR

The TruPDF OCR engine — WASM, runs fully in-browser. ~100 languages. Lazy-loaded (~6 MB compressed) so unused sessions pay zero.

Recognize

import { OCREngine } from '@trupdf/viewer';
 
const ocr = new OCREngine({
  languages: ['eng', 'fra'], // first is primary, rest are fallbacks
  onProgress: ({ page, status, progress }) => console.log(page, status, progress),
});
 
await ocr.init(); // boots the worker + loads language models (idempotent)
const result = await ocr.recognize(canvas, 1); // (image, 1-indexed page)
 
for (const line of result.lines) {
  console.log(line.text, line.confidence, line.box);
}
 
await ocr.destroy(); // terminates the worker, frees WASM memory

Inputs accepted: HTMLCanvasElement, HTMLImageElement, Blob. Rasterise a page first (e.g. via ExportEngine.exportImage) before calling recognize. Concurrent recognize calls are serialised internally, so it is safe to fire one per page in a loop.

`OCRPageResult`

Field	Meaning
`page`	The page number you passed in
`text`	Full recognised text (trimmed)
`confidence`	Average confidence across all words (0–100)
`lines`	`OCRLine[]` — `{ text, confidence, box, words }` per line
`width`	Source image width in pixels
`height`	Source image height in pixels
`durationMs`	Wall-clock recognition time

Line and word box values are pixel-space bounding boxes ({ x, y, width, height }, top-left origin) so you can paint highlights straight onto the source canvas.

Languages

The OCRLanguage type covers the common codes:

eng (English), fra (French), deu (German), spa (Spanish), ita (Italian), por (Portuguese), rus (Russian), chi_sim (Chinese, simplified), chi_tra (traditional), jpn (Japanese), kor (Korean), ara (Arabic), heb (Hebrew), hin (Hindi), tha (Thai), vie (Vietnamese), tur (Turkish)

Language models are fetched on demand — only the languages you list are downloaded, on first init().

Self-hosting the engine assets

Air-gapped or CDN-restricted deployments can serve the OCR engine's worker and core assets — plus the language models — from their own origin and point the engine at them:

const ocr = new OCREngine({
  workerPath: '/static/ocr-worker.js',
  corePath: '/static/ocr-core.wasm.js',
  langPath: '/static/lang/',
});

workerPath and corePath override where the engine's worker script and WASM core load from; langPath is the URL prefix for the per-language model files. Serve the files yourself and every OCR request stays inside your network.

From the viewer

When aiConfig is set with OCR fallback, the AI panel auto-OCRs scanned pages before the AI pipeline runs. You can also surface OCR as its own toolbar action and re-flow recognised text as a selectable text layer (standard mode only).

License gating

OCREngine calls gate.require('ocr') in its constructor. Trial tokens cannot OCR.