datalab-to/indra is an OCR engine that turns PDFs and images on the CPU into Markdown, plain text, or JSON that carries box coordinates. According to the repository description, the model is on the order of one million parameters, and the latest release is v1.1.2. The information provided contains no changelog for this release. It is written in Python and released under the Apache-2.0 license.
It finds text lines in a document, fixes the reading order, and then outputs the recognition result. CTC is an approach that recognizes strings instead of assigning a position to each character. The repository states that it processes 49 pages per second on CPU, describing it as 15 times faster than Tesseract with accuracy close to models that have more than a billion parameters. That comparison rests on the repository's own benchmark. Alongside PDFs and images, it says it handles handwriting, more than 80 languages, rotated or skewed scans, document layout, and the text layer of digital PDFs. It also provides a command-line tool and a local web viewer, where lines, blocks, and the reading order can be checked. These features can be used in pipelines that convert documents into multiple formats.
The performance comparison is not an independently verified result but a measurement taken by the project itself. The repository lists handwriting handling and cases where Korean lines are read as Hanja among its failure points. It explains that the best-in-class models sit behind commercial platforms and cannot be checked in code. The presented figures should therefore be read as reference material, and the recognition results on real documents need to be verified separately.