docext (Nanonets-OCR)

Nanonets open-sourced docext as an on-premises document intelligence toolkit built around its Nanonets-OCR-s model, a compact 3B parameter vision language model trained for image-to-markdown conversion. The conversion path handles PDFs and images with LaTeX equation recognition, table extraction, signature and watermark detection, and checkboxes rendered as Unicode symbols, while a separate OCR-free extraction path pulls key fields and tables directly from documents like invoices and passports with confidence scores attached, so low-confidence values can be routed to human review. Multi-page processing, a REST API, and a benchmarking leaderboard covering several document intelligence tasks round out the package. It installs from PyPI as docext with a Dockerfile for containerized deployment on Linux and macOS, and running the bundled model locally effectively requires a GPU. The repository LICENSE is Apache-2.0, it has about 2,000 stars, and the model weights are on Hugging Face. Teams pick it when document data cannot leave their infrastructure but they still want VLM-grade extraction quality.