A document parser that keeps your files local—until it can't

What it does

LiteParse is a Rust-core document parser that turns PDFs, Office documents, and images into structured text or JSON with precise bounding boxes. It bundles Tesseract for zero-setup OCR, can proxy to external HTTP OCR services like EasyOCR or PaddleOCR, and renders page screenshots for LLM agents. Everything runs locally across Linux, macOS, Windows, and even the browser via WASM, with language bindings for Python, Node.js, and TypeScript.

The interesting bit

The README is unusually candid: it explicitly warns that dense tables, multi-column layouts, charts, and scanned PDFs will likely outperform the local tool and points you toward the vendor’s cloud parser, LlamaParse. That honesty is rarer than it should be in open-source tooling. The architecture is also pleasantly modular—PDFium handles text extraction, LibreOffice and ImageMagick handle format conversion, and the OCR layer is swappable via a simple HTTP API spec.

Key highlights

• Runs fully offline with bundled Tesseract OCR, but accepts plug-in HTTP OCR servers for heavier workloads
• Outputs plain text or structured JSON with spatial bounding boxes for every extracted element
• Generates high-DPI page screenshots specifically designed for LLM agent consumption
• Supports automatic conversion of Word, Excel, PowerPoint, and image files to PDF before parsing
• Ships identical lit CLI across Rust, Python, and Node.js installations, plus a WASM build for browsers

Caveats

• Complex documents—dense tables, handwritten text, scanned PDFs—are explicitly called out as weak spots compared to cloud-based alternatives
• Office and image format conversions require external system dependencies (LibreOffice and ImageMagick) that are not bundled
• The README is truncated in the provided source, so full performance characteristics and memory usage are unclear

Verdict

Developers building local-first document pipelines or LLM agent skills who can tolerate occasional OCR imperfections should look here. If you need guaranteed accuracy on gnarly scanned invoices or multi-column academic papers, the tool itself suggests you look elsewhere.