LlamaIndex has announced a complete rewrite of its open-source document parsing library LiteParse, releasing version 2.0 with a pure Rust core. According to the project, the new parser is up to 100x faster when handling small documents, and nearly 3x faster on larger ones. The move is designed to provide a local, lightning-fast layout parsing foundation for AI agents and retrieval-augmented generation (RAG) pipelines — no large language model required.
LiteParse 2.0 maintains its no-LLM, fully local design. It integrates a deeply customized fork of PDFium for spatial layout analysis and uses the tesseract-rs library for local optical character recognition (OCR). The tool currently supports PDF and Office documents (DOCX, XLSX, PPTX). The parser projects text into a two-dimensional spatial layout, outputting structured text that preserves positional and layout relationships — giving LLMs high-fidelity context with minimal power consumption.
On the ecosystem side, LlamaIndex provides native package support across major runtimes. Developers can get started via Python (pip install liteparse), JavaScript (npm i @llamaindex/liteparse), or Rust's Cargo registry. Thanks to the Rust base, the new version compiles to WebAssembly, enabling local execution in browsers and edge computing nodes. However, in WebAssembly environments, OCR functionality is not built-in — developers need to inject an external callback (e.g., via tesseract.js) to handle document scanning.