Sparrow is an open-source document intelligence toolkit from KatanaML that extracts structured data from invoices, receipts, forms, statements and other documents using large language models and vision LLMs. You define the fields you want as a JSON schema, and Sparrow returns matching, validated structured data — no template building or rules engines. The GPL-3.0 project has more than 5,200 stars on GitHub and is under active development in 2026.
Sparrow combines several extraction strategies behind one interface: instruction calling to LLMs, vision models that read document images directly, and an agentic layer — Sparrow Agent — that chains classification, extraction and field validation with visual monitoring and error handling. Business rules such as formatting, derived fields and classification can be expressed as typed schema fields and pushed to the model itself, cutting the post-processing code that document pipelines usually accumulate. It supports multi-page PDFs, page classification and bounding-box annotation, validates extracted values against your schema types before returning them, and runs locally — with Ollama models for full privacy or cloud LLMs when you prefer — with no cloud dependency required. The hosted demo at sparrow.katanaml.io shows the workflow live, and KatanaML offers to set Sparrow up on company infrastructure, which together make it accessible both to developers self-hosting the open-source version and to businesses that want a supported deployment.
Freemium. Sparrow is free and open source (GPL-3.0) to self-host; KatanaML offers hosted access keys and paid on-infrastructure setup and pilots for businesses.
Developers and operations teams who need reliable structured data out of documents and want an open-source, locally runnable alternative to proprietary document AI services.