blazegraph-io

Parse PDFs into semantic document graphs with bounding boxes. Built for GraphRAG.

55 pages  →  3,022 text elements  →  94 nodes  →  1.1s

Install

cargo install blazegraph-io

No account needed. No API key. Runs entirely on your machine.

On first run, the CLI downloads a Java Runtime (~60MB) for PDF text extraction. It's cached for future use.

Usage

# Parse a PDF, output to stdout
blazegraph-io parse document.pdf

# Write to a file
blazegraph-io parse document.pdf -o bgraph.json

# Use a custom config
blazegraph-io parse contract.pdf -c my-config.yaml -o bgraph.json

What You Get

Every node in the output graph has:

• Semantic location — tree position (path: "2.3"), depth, breadcrumbs
• Physical location — page number, bounding box (x, y, width, height in PDF points)
• Content — the node's text with pre-calculated token count
• Relationships — parent and children UUIDs for tree navigation

{
  "node_type": "Paragraph",
  "location": {
    "semantic": { "path": "2.2", "depth": 2, "breadcrumbs": ["paper.pdf", "Introduction"] },
    "physical": { "page": 1, "bounding_box": { "x": 91.9, "y": 585.9, "width": 427.5, "height": 164.2 } }
  },
  "content": { "text": "The fundamental problem of communication..." },
  "token_count": 206
}

This dual location model is what makes the output GraphRAG-ready: ground LLM outputs to specific physical locations in the original PDF.

Configuration

The default config works well for most documents. For specific document types, create a YAML config file and reuse it across similar documents.

See the Configuration Reference for all tuning parameters.

As a library

If you want to embed the parser in your own Rust application, use blazegraph-io-core instead.

Python SDK

A typed Python SDK is also available:

pip install blazegraph-io

See the Python SDK Guide.

Hosted API

Same parser, no infrastructure. Available at blazegraph.io. 500 free credits on signup.

License

Licensed under either of:

• Apache License, Version 2.0
• MIT license

at your option.