Drop an image or PDF to convert

Code: Apache-2.0. Model weights: Datalab's, under a modified OpenRAIL-M license with use restrictions (revenue and funding limits, no competing product). Read the license status and the model card before you use them. Not endorsed by Datalab.

Configuration

The defaults work well for most files. Adjust only what you need, then apply your changes.

Pipeline Image engine and what ends up in the markdown

Image engine opencv downloads OpenCV.js (about 10 MB) the first time you apply it.

Read math as LaTeX (loads four more models, about 20 MB) Keep page headers and footers in the markdown Second detection pass on pages with small text

Models Datalab's weights, cached in this browser

Model set v1.1 adds handwriting, 80+ languages, tilted and vertical pages (about 30 MB). v3.1 reads Latin, Cyrillic, Greek and Chinese (about 18 MB).

Re-read unsure lines with the language models (v1.1, 46 MB more) Warm up after applying (hidden run so the first convert is fast)

Models stay in this browser after the first download, so a reload does not fetch them again.

The weights are datalab-to/indra-ocr. Loading them means you accept their license.

PDF documents Applies to the next convert, no reload

Text source

Maximum pages Pages are read one after the other.

Render resolution (dpi) Used when a page is read with OCR.

Password

Runtime & debugging Execution provider and console logs

Execution provider Automatic is the tested path. WebGPU may fail or give different results.

Show detailed logs in the browser console

Using recommended defaults

Page Input

Upload an image or a PDF to begin

Sample files

Markdown

// The markdown will appear here

Plain text

// One line per row, in reading order

JSON Output

// Inference output will appear here

ppu-indra-ocr

Preparing...

This downloads Datalab's Indra OCR weights (about 18 MB), which have their own license: a modified OpenRAIL-M with use restrictions. See the license status and the model card. Not endorsed by Datalab.