How Chandra OCR 2 Accurately Parses Complex Tables and Handwritten Text

Beyond Simple Recognition: Solving Traditional OCR Pain Points

Traditional OCR struggles with complex tables, handwritten text, mathematical formulas, and mixed‑layout documents, often losing structural information.

Chandra OCR 2 adopts a “document intelligence” approach, building a deep‑understanding model of page elements and their relationships.

Core Technical Highlights

Layout preservation : precisely reproduces tables, columns, heading hierarchies, and lists.

Complex element handling : strong support for mathematical formulas, handwritten text, and form checkboxes.

Multilingual coverage : supports 90+ languages and performs well on mixed‑language documents.

Structured output : generates HTML or JSON directly for downstream workflows.

Result: a scholarly paper with a complex table is output as a clean Markdown table; a handwritten application form yields both text and checkbox states.

Performance: Benchmark Results

The project released detailed benchmark data. In the “olmocr” comprehensive benchmark, Chandra 2’s scores substantially exceed those of other open‑source and commercial models.

Because public multilingual OCR test sets are scarce, the team created a custom suite covering tables, formulas, layout, and text accuracy across languages such as Chinese, Japanese, and Arabic. The suite shows consistently high precision for all 90+ languages.

“Multilingual performance is a key focus of Chandra 2. We built a benchmark covering tables, math, sequence, layout and text accuracy, and the results demonstrate robust performance on over 90 languages.” – project team

Rapid Onboarding

Developers can install the package via pip and run either a local Hugging Face inference mode or the more efficient vLLM server mode.

pip install chandra-ocr
# start vLLM service (recommended, lightweight and efficient)
chandra_vllm
# convert a document
chandra input.pdf ./output

A free online Playground and a hosted API are also provided for immediate experimentation.

Application Scenarios

Education & research : digitization of historical archives and scientific papers, especially those containing many formulas.

Finance & legal : automatic processing of scanned financial statements and contracts to extract structured data.

Office automation : bulk conversion of scanned forms and applications into queryable databases.

Content publishing : transformation of legacy books and magazines into re‑flowable electronic formats.

Open Source and License

The code is released under the Apache 2.0 license; the model uses the OpenRAIL‑M license, enabling both open‑source collaboration and commercial use. An active Discord community supports developers.