C CreatorPrivacyKit

Read a document without sending the document

Local OCR is a Documents job that happens to use a model. The product question is “can I get text and a searchable PDF without an upload queue?” — not “which cloud model is smartest”.

Reviewed 17 September 2026

What it is for

Languages

Shipped recognition packs: English, Simplified Chinese, Traditional Chinese, Japanese. You choose a language before running. Auto does not detect script; it falls back to English. If you leave Auto on a Chinese scan, expect English-biased garbage rather than magic.

Privacy and limits

Recognition runs in the tab after you start it. Runtime assets load from this origin, not from a third-party OCR API. Encrypted PDFs are rejected; we do not bypass passwords.

Accuracy drops on heavy JPEG compression, skewed pages, tiny UI type, handwriting, and mixed receipts. We do not advertise a single accuracy percentage.

Searchable PDF overlay uses Tesseract word boxes. CJK export needs the bundled font. Force OCR on a digital PDF may stack an OCR layer on top of the original text layer — that limitation is disclosed in the tool.

Related

Open Local OCR · Image to text without uploading · Searchable scanned PDFs · Receipt OCR failure modes · Methodology.