Help, privacy & practical limits
Your source images and extracted text stay in session memory. The hosting provider still serves ordinary app, font and OCR asset requests.
Local processing boundary
No image/text uploads, telemetry, analytics, accounts, cloud fallback or document persistence are configured. Source pixels are sent only to the local worker. Cache Storage holds only known OCR engine/language files. This cache is separate from your documents and can be removed independently. Original files are not overwritten.
Resource and format limits
Up to 10 images, each at most 25 MB / 50 MP / 12,000 px per source edge. Working images are downsampled to 10 MP / 5,000 px per edge. A session is limited to 40 MP total working pixels and 100 MB normalized images. Text is capped at 1 million characters per image. These provisional guardrails cannot guarantee capacity on low-memory devices. Direct PDF input, searchable PDF, tables and automatic language detection are outside this MVP.
Cancel, clear and recover
Cancel terminates the native OCR worker, including during initialization, and aborts in-flight asset downloads; completed images remain. Native browser image decoding/encoding can finish after cancellation, but late results are discarded. Clear workspace and pagehide release document references, previews and workers; this is not secure erasure of browser/OS memory. Export before leaving.
Compatibility
Image import needs createImageBitmap and Canvas. Local OCR needs Web Workers and WebAssembly. Asset verification needs Web Crypto on HTTPS/localhost. Offline OCR additionally needs Service Worker and Cache Storage. Without those caching APIs, online local OCR can still run after verified downloads. If clipboard write fails, manually select/copy or download TXT.
Known verification limits
Automated tests use installed desktop Chrome and synthetic printed-text fixtures. Mobile hardware, real photographed documents, Firefox, Safari, Edge and screen-reader sessions require separate validation. Low confidence or empty text calls for a sharper image, correct language and better crop; high confidence can still contain errors. TXT is UTF-8; JSON preserves raw/edited Unicode text and source order.