datalab-to/indra

1M param OCR model - runs on CPU, 15x faster than tesseract, close to 1B+ param models in accuracy

What is Indra OCR?

Indra OCR is a tiny (≈1 M‑parameter) optical‑character‑recognition model that runs on CPUs. It can read PDFs and images at up to 49 pages / second (≈230 ms per page on 4 x86 vCPUs) and outputs the result as markdown, plain text, or JSON with detailed layout information.

Key Features

Speed – 11‑17× faster than Tesseract/PaddleOCR on the same hardware.
Accuracy – OmniParseBench headline score 62.1, beating PaddleOCR’s full pipeline (58.4).
Layout awareness – bounding boxes for every line and block, block labels (headings, tables, captions, footnotes, etc.) and reading order.
Hybrid PDF handling – uses a PDF’s native text layer when it passes quality checks; otherwise falls back to OCR, saving work.
Small footprint – ~18 MB ONNX model, no PyTorch, runs with ONNX Runtime, OpenCV and pypdfium2.
Multi‑language OCR – Latin, Cyrillic, Greek, Devanagari, Bengali, Chinese (simplified & traditional). Digital PDFs support any script via the text layer.
Optional math support – --math flag adds a classifier and a second recognizer that outputs LaTeX‑style inline math.
Local viewer – indra-ocr-viewer shows pages with overlayed boxes, markdown, JSON and timings.

Quick Start (CLI)

pip install indra-ocr               # install the package
indra-ocr paper.pdf                # markdown to stdout
indra-ocr *.pdf --out results/     # one .md per file, all cores
indra-ocr scan.png --format json   # JSON with lines, boxes, confidences
indra-ocr-viewer                   # open http://localhost:8765

The first run downloads the ONNX model from Hugging Face (≈18 MB).

Python API (basic example)

import indra_ocr
conv = indra_ocr.load(threads=4)          # auto mode (text‑layer + OCR)
for i, page in conv.convert("paper.pdf"):
    print(page.markdown)                   # markdown with headings, tables, etc.
    print(page.blocks)                     # list of layout blocks with labels

For bulk conversion use indra_ocr.convert_many([...], cores=None) which spawns one worker per CPU core.

Output structure

markdown – formatted text (headings, lists, tables as HTML). Headers/footers omitted unless --keep-furniture.
text – plain‑text, one line per detected line.
lines – list of Line(box, text, confidence) objects.
blocks – layout blocks with label, bbox, text in reading order.
source – text (PDF layer), ocr, hybrid or error.
timings – per‑stage milliseconds, plus total.
Additional meta like rotation, skew, and a short note explaining the processing path.

When to use Indra OCR vs. the managed service

Indra OCR shines for clean printed documents where CPU‑only speed is essential. It is not suited for:

Complex tables with merged cells or spanning rows.
Handwritten text.
Scripts outside the OCR list (Arabic, Hebrew, Thai, Japanese, Korean) – those work only via the PDF text layer.
Very degraded or low‑resolution scans.
Vertical text, banner‑wide pages, engineering‑drawing labels, or pages > 100 MP at 200 dpi. For those cases the authors recommend their higher‑accuracy, GPU‑backed model Chandra via the Datalab managed platform.

Licensing & Commercial Use

Code – Apache 2.0.
Model weights – Modified OpenRAIL‑M (free for research, personal use, and startups < $20 M revenue). Commercial use that competes with Datalab’s API requires a separate license (see the pricing page).

Installation notes

pip install indra-ocr                     # stable release
# or latest GitHub version
pip install git+https://github.com/datalab-to/indra.git
# optional Tesseract comparison support
pip install "indra-ocr[tesseract]"

Supported on Python 3.10+, Linux (x86_64 & aarch64) and macOS (Apple silicon). The package pulls ONNX Runtime, OpenCV and pypdfium2 automatically.

Benchmarks at a glance

Metric	Indra OCR (default)	PaddleOCR PP‑OCRv6 (text only)	Tesseract 5
Throughput (pages / s) – 32 cores	49.4	4.28	2.92
Median latency per page	228 ms	2 685 ms	5 496 ms
OmniParseBench headline	62.1	–	–
Text score	41.3	35.8	19.3
Table score	56.4	0.0	0.0
Layout score	88.6	–	–

Where to learn more

Homepage: https://www.datalab.to
Model on Hugging Face: https://huggingface.co/datalab-to/indra-ocr
Docs & Playground: https://documentation.datalab.to & https://www.datalab.to/playground
Discord community: https://discord.gg/KuZwXNGnfH

TL;DR

Indra OCR is a lightweight, CPU‑only OCR engine that delivers high speed and respectable accuracy for printed documents, with rich layout output and an easy‑to‑use CLI/Python API. It’s ideal for batch processing of clean PDFs or images when you need results fast and can tolerate its limitations on complex tables, handwriting, or non‑Latin scripts.