FAQ
Learn about accuracy limits, experimental Korean handwriting, downloads and storage, PDFs and word tools.
How accurate is OCR?
It is suited to clear printed text, but no accuracy or correct-answer percentage is guaranteed. Blur, small text, handwriting, vertical writing and complex tables can cause errors. Engine confidence is not actual accuracy; compare and edit against the source.
Can it read Korean handwriting?
Korean line recognition is experimental. Printed OCR first finds lines, then a Korean model reads them. Real handwriting may have low accuracy or no detected lines. English handwriting mode is not for Korean.
How large are the models and where are they stored?
Settings show engine and language sizes and the amount to download. English handwriting needs about 63.6 MB of q8 weights plus configuration, tokenizer and runtime files. The Korean model and selected print languages download separately. Cache Storage holds downloads; use Delete in settings to remove them. Cleared data must be downloaded again.
Are images, PDFs or results sent to a server?
No. Recognition and PDF processing happen inside your browser, without uploading files or results. Downloading software, models and language data uses the network. The English handwriting model downloads from Hugging Face.
Can I extract only some PDF pages?
Yes. After adding the file, enter a range such as 1-3,5. Leave it empty for all pages. Text layers are extracted directly; pages without one use OCR. You can copy all result pages or just the selected page.
Which languages are supported?
Printed OCR offers Korean, English, Japanese, Simplified and Traditional Chinese, Spanish, Portuguese, French, German, Vietnamese, Indonesian, Turkish, Italian and Polish. You can combine languages. English handwriting and Korean line recognition are separate experimental modes. Translation is not provided.
Can word tools copy only particular codes?
After recognition, use starts with, contains, ends with or pattern filters. The easy pattern W+digits matches whole words such as W123. Choose deduplication and a separator, then copy. The list uses original recognized words before editing.
Does it work offline?
Processing may work without a network in an already open tab when all required code, engines, languages and models are ready. Offline reopening and every feature are not guaranteed. First-time features and cleared data need downloads.