Chandra OCR

Why it's ranked here

The strongest case is breadth plus deployability. One pipeline covers PDFs, common image formats, directories, page selection, figures, forms, and layout-aware text. The published comparison reports strong results across tables and difficult scans. However, most comparison figures are the project’s own benchmarks, so validate representative documents before committing.

What's good

The output pipeline does more than copy model text. It filters blank pages, optionally removes headers and footers, repairs missing paragraph structure, retains HTML tables inside Markdown, preserves math delimiters, and crops figure regions into separate images. Batch controls and retry logic make the remote inference path practical for larger jobs.

Tradeoffs

Local inference pulls in Torch and Transformers, while the supplied vLLM launcher assumes Docker, NVIDIA GPUs, and elevated privileges. The managed service is described as more accurate than the open weights. Code uses Apache 2.0, but model weights carry commercial restrictions, including limits tied to company size and competition with the vendor API.

How to use it well

Use it for document conversion pipelines where page structure, formulas, tables, or extracted figures matter more than plain text alone. Start with the interactive app on representative pages, then move repeatable work to directory batches through vLLM. Choose local Hugging Face inference when self-contained execution matters. It does not supply the model runtime hardware for you.

Observed

Code license

Apache 2.0

Model license

Modified OpenRAIL-M with stated commercial-use restrictions

Primary language

Python

Runtime requirement

Python 3.10 or newer

Installation surface

pip package with optional Hugging Face, Streamlit app, and combined extras

Interfaces

Command-line tools, Streamlit application, local Hugging Face inference, and OpenAI-compatible vLLM inference

Input surface

Single files or directories containing PDFs and common raster image formats

Output surface

Markdown, HTML, JSON metadata, layout chunks, and extracted images