Sparrow

Katana ML's Sparrow approaches document extraction as an API-first pipeline: documents like invoices, receipts, bank statements, and tables go in, validated structured JSON comes out, driven by vision language models and agentic multi-step workflows. Backends are pluggable, with MLX for Apple Silicon, vLLM for NVIDIA hardware, Ollama, Hugging Face cloud GPUs, and the Mistral OCR cloud service all supported, so the same pipeline can run on a MacBook, a GPU server, or against a hosted model. A FastAPI layer exposes inference and instruction-inference REST endpoints with Swagger documentation, and query-style instructions select which fields to extract from PNG and JPG images or multi-page PDFs. Setup is more involved than a single pip install: Python 3.12.10 or newer via pyenv, separate virtual environments per pipeline component, and Poppler for PDF handling. The license is GPL-3.0 with dual licensing available, free for open-source projects and organizations under 5 million dollars in revenue. With about 5,200 stars, it serves teams standing up private document extraction services without training custom models.