Prompt & Done · BOTTLESCAN pack

A crooked photo in, clean markdown out. Here's the install command.

Indra OCR (datalab-to/indra) is a small open-source OCR model that runs on a CPU: no GPU, no PyTorch, files stay on your machine. Python 3.10 or newer.

1 Install

pip install indra-ocr

The first run downloads the models (about 76 MB) from Hugging Face and caches them.

2 Point it at a photo or PDF

indra-ocr photo.jpg
indra-ocr scan.pdf                    # markdown to stdout
indra-ocr *.pdf --out results/        # one .md per file
indra-ocr photo.jpg --format json     # lines, boxes, layout blocks
indra-ocr photo.jpg --timings         # per-page time on stderr

Markdown prints to your terminal. indra-ocr-viewer opens a local viewer on http://localhost:8765.

3 What we measured

Tested 26 Oct 2026 with indra-ocr 1.1.2 on an Apple M5 Max (CPU only). A crooked test photo of a printed order page (heading, paragraph, a 3-column table) came out with the heading, text, every table row and every price correct: about 0.5 s for the command, 73 ms of OCR for the page. The same page rotated 180 and 90 degrees gave byte-identical markdown. A receipt photo got every price right but merged the address into the heading and turned the first item into a table header; a skewed scanned PDF came out exact.

Fair to say: our test photo was made for the demo (a rendered page warped to look like a phone photo), printed text only. The README rates handwriting as rough (25.2% on its handwriting tests), and cursive is the worst. Code is Apache-2.0; the model weights are modified OpenRAIL-M (free for research, personal use and startups under $20M, not for competing with Datalab's API).