Open OCR → and choose “Try a text sample”. OmniKit creates a 1200 × 500 pixel image locally and selects English recognition. The practice image contains no customer data.
Keep the expected transcription beside the result
OmniKit sample Invoice 2026-001 Total 123.45
Start recognition and wait for the result. First use downloads the engine and language model; the image itself is processed on your device. A slow connection can delay the first run. Retry if loading fails.
Check meaning, not just completion
| Field | Expected | What to inspect |
|---|---|---|
| Identifier | 2026-001 | Keep the hyphen and leading zeros. Distinguish zero from O. |
| Amount | 123.45 | A missing decimal changes the value. Do not silently interpret a space as a decimal. |
| Order | Title, identifier, amount | Columns and tables can be read in an unexpected order. |
Edit the result before copying or downloading TXT. Average confidence is an engine estimate, not proof that an identifier or amount is correct. Compare important fields character by character.
When OCR is unnecessary
First try selecting and copying text in the original PDF. With “Prefer embedded PDF text” enabled, the tool uses an available text layer directly. This avoids recognizing text again, but a PDF can still have incorrect character encoding or reading order. Image recognition is needed for scanned pages without useful embedded text.
Improve a difficult result
- Match the recognition language to the source; choose Chinese + English for mixed text.
- Use upright, sufficiently large text on a clean background. Avoid strong compression first.
- Compare the original with enhancement modes. Binarization can erase faint text.
- Manually review handwriting, tables, skewed pages, and complex layouts. TXT does not preserve table formatting.
The batch limit is 20 files and 20 pages; each file may be up to 30 MB. Cancellation stops recognition. Review any partial text before using it. Refreshing does not restore the workspace, so copy or download your result.