OCR Evidence Workbench
What OCR Evidence Workbench actually does
A successful-looking text box is not proof of a faithful transcription. This workbench runs a pinned Tesseract LSTM engine locally, keeps every observed pass, chooses one whole observed pass per page instead of inventing unseen text, and lets you correct the editable result without rewriting the raw evidence.
How to use it
- Choose one PNG, JPEG, WebP, BMP, or PDF and select English, Spanish, or both languages.
- Run quick or redundant OCR; PDF pages are rendered locally, every scheduled pass retains its status, configuration, and warnings, and every valid completed pass retains text, TSV geometry, and hOCR.
- Review low-confidence words beside the source, edit the complete result, then export current text, canonical JSON, or a raw evidence ZIP.
Useful for
- Recover English or Spanish text from screenshots, scans, photographed pages, and image-only PDFs.
- Inspect punctuation, accents, code-like spacing, and ambiguous words instead of hiding disagreement behind one output.
- Keep an auditable engine record separate from later human corrections.
Limits worth knowing
- Browser input is capped at 25 MiB, 12 PDF pages, 12 million pixels per page, 48 million rendered pixels, 36 OCR passes, and 64 MiB of uncompressed evidence; balanced mode therefore accepts at most nine pages.
- Handwriting, equations, arbitrary scripts, exact reading order, tables, typography, bold or italic reconstruction, and universal font fidelity are not promised; review the source and uncertain tokens.
- The browser surface does not create searchable PDFs or ALTO and is intentionally narrower than the local CLI evidence pipeline.
Questions people ask
Does the image or recognized text leave this device?
No. The pinned worker, OCR core, English and Spanish models, and PDF support files are served as static same-origin assets; file bytes and recognized text stay in this tab and no OCR cache is persisted.
Is the confidence score a probability that a word is correct?
No. It is raw engine evidence used to prioritize review, not a calibrated probability or a guarantee.
Why preserve several passes?
Different preprocessing and page-segmentation assumptions expose different mistakes. The workbench retains those observations, while the canonical page remains one traceable engine pass.