Schema-driven extraction for documents that refuse to cooperate

What it does Documind is a Node.js library that feeds documents—PDFs, Word files, images, HTML, or plain text—to an LLM and returns structured JSON. You define a schema upfront (field names, types like string, number, boolean, enum, plus nested arrays and objects) or pick from built-in templates for common formats like invoices and bank statements. It also converts documents to Markdown and can auto-generate schemas from the content itself.

The interesting bit The project is essentially a schema-enforcing orchestration layer over Zerox and various LLMs. What keeps it from being a thin wrapper is the strict contract: instead of dumping raw OCR text into a prompt and hoping, you declare a typed schema and get back shaped JSON. It also runs locally via Llama3.2-vision or Llava if you’d rather not ship sensitive documents to OpenAI.

Key highlights

• Supports custom nested schemas with typed fields including boolean and enum
• Built-in templates for common documents such as bank statements and invoices
• Multi-format input: PDF, DOCX, PNG, JPG, TXT, HTML
• Local LLM support via Llama3.2-vision and Llava for fully offline extraction
• Auto-generates schemas from document content when you’d rather not hand-write one
• AGPL v3.0 licensed, built on top of Zerox

Caveats

• Requires native system dependencies (Ghostscript and GraphicsMagick), so it is not a pure-JS install
• Node.js 18+ required
• Several roadmap items—image extraction, fine-tuned models, advanced formatters—are not yet released

Verdict A solid choice for developers who want schema-guaranteed document extraction without building an ML pipeline. Less appealing if you’re allergic to installing Ghostscript and GraphicsMagick, or if you need the upcoming image-extraction and fine-tuned model features that are still on the roadmap.