indra: Fast CPU OCR with Datalab’s Indra Model

Overview

Indra is Datalab’s CPU-focused OCR tool, built around its Indra OCR model and ONNX Runtime. It converts PDFs and images into markdown, plain text, or JSON, addressing the need for fast document extraction with line boxes, layout blocks, and reading order rather than text alone. It supports Linux and macOS, and can use a PDF’s digital text layer as well as OCR scanned pages.

The project is aimed at clean printed documents and high-throughput CPU processing. It is less suited to handwriting, degraded scans, complex tables, or scanned text in scripts outside its supported OCR set.

Key Features

• Converts PDFs and images to markdown, plain text, or JSON through a CLI or Python API.
• Uses a small ONNX Runtime model package and does not require PyTorch or a GPU.
• Combines a PDF’s text layer with OCR when needed, including OCR for text inside images.
• Returns line-level boxes and confidences, layout blocks, reading order, page source, and timings.
• Labels layout elements such as headings, tables, captions, lists, and page furniture.
• Provides a local web viewer for inspecting page overlays and extracted output.
• Supports OCR for Latin, Cyrillic, Greek, Devanagari, Bengali, and Chinese, while digital PDF text extraction supports any script.

Use Cases

• Teams processing large batches of clean, printed PDFs on CPU can use the CLI or Python interface to extract markdown without deploying GPU infrastructure.
• Developers building document-ingestion pipelines can use JSON output to retain line coordinates, layout labels, and confidence data.
• Researchers or analysts handling digital PDFs can use hybrid mode to preserve text-layer content while OCRing missing text in page images.
• Users reviewing OCR output can inspect line and block overlays in the local viewer before using extracted content downstream.

Getting Started

pip install indra-ocr

The first run downloads the models. See the README for CLI, Python, viewer, and configuration details.

Alternatives

• marker: Marker converts PDFs and other documents to Markdown, JSON, HTML, or chunks, using local models and an inference backend.
• docling: Docling handles formats beyond PDFs and images, including Word files and spreadsheets, and offers a Python library and CLI.
• PaddleOCR: PaddleOCR is a broader OCR toolkit with multilingual recognition and document-layout parsing for images and PDFs.
• pdf-craft: PDF Craft focuses on scanned books and documents, with editable Markdown or EPUB output and optional translation.

Considerations

The repository is only a few days old in the supplied project-health data, with a recent release and pull requests but little commit history to establish maintenance patterns. The measured contributor data shows that all contributions come from one person, and no issues were closed in the measured period, so teams should assess their support and continuity needs.

OCR is optimized for clean print. The README identifies handwriting, degraded scans, complex tables, vertical text, and several scripts on scanned pages as limitations. The code is Apache-2.0, but the model weights use a modified OpenRAIL-M license with additional usage terms, including restrictions on competitive use against Datalab’s API. Check both licenses before adopting it commercially.