Chandra OCR

Chandra OCR converts document images and PDFs into text that retains page structure. It's for people processing scans, handwritten records or documents whose tables and forms need to survive extraction.

• Structured output: Export Markdown, HTML or JSON with layout information for documents with multiple columns, tables and mathematical notation.
• Handwriting and forms: Recognize handwritten text and reconstruct forms, including checkboxes.
• Multilingual text: Process languages including Arabic, Chinese, Hindi, Japanese and Ukrainian.
• Images and diagrams: Extract visual content alongside the text, with captions and structured data.
• Batch processing: Process individual files or whole directories, or use a Streamlit interface for single-page processing.

The model runs locally through HuggingFace or on a self-hosted vLLM server, with Docker support for the server. Local processing keeps documents on your hardware; a remote backend processes them on the server you connect to. The vLLM backend can use NVIDIA GPUs.

The code is open source under Apache 2.0. Model weights have a separate modified OpenRAIL-M license, with free research and personal use, restrictions on commercial use, and a prohibition on competing with Datalab's API. Commercial use depends on the model license's eligibility limits and restrictions.

Datalab also offers a hosted service with a different, higher-accuracy model and managed batch processing. Documents sent to that service leave your machine; the service uses zero data retention by default.