Self-hosted document parsing with a shopping list of local backends

What it does

Sparrow is an API-first system for enterprise document intelligence. You feed it PDFs or images—bank statements, invoices, tables—and it returns validated JSON via REST. It also orchestrates multi-step agent workflows and instruction-based text processing, all designed to run on your own hardware.

The interesting bit

Instead of locking you into one inference stack, Sparrow treats backends like interchangeable lenses: MLX on Apple Silicon, vLLM on NVIDIA, Ollama, or Hugging Face, all exposing the same API surface. That flexibility is the product’s core bet—local deployment with hardware-agnostic plumbing.

Key highlights

• Runs fully offline with local Vision LLMs (Qwen, Mistral, DeepSeek OCR, Gemma 4) and text models
• Pluggable pipeline architecture mixing Vision LLM (sparrow-parse), Text LLM (sparrow-instructor), and agentic steps
• JSON schema validation baked into the extraction output
• Visual workflow monitoring via Prefect for agent orchestration
• Enterprise trimmings: rate limiting, usage analytics, and commercial licensing available

Caveats

• Setup is platform-picky: separate Python virtual environments per pipeline, manual poppler installs for PDF support, and GPU memory requirements that vary by model
• The README warns you to check platform-specific dependency references before installing, which suggests the out-of-box experience is not quite turnkey

Verdict

Worth evaluating if you need structured document extraction in a regulated or air-gapped environment. Skip it if you are looking for a managed SaaS with zero infrastructure overhead.