Documents & data

run-llama/liteparse

Local, open-source document parser, no API or GPU

12k
Apache-2.0
Rust

What it is and why it matters

liteparse is LlamaIndex's open-source document parser, a lighter version of their Llama Parse product. It is TypeScript-native (with a Python wrapper), builds on PDF.js and Tesseract.js, needs no API key or GPU and supports around 50 formats, including images. Instead of detecting tables and converting them, it preserves spatial layout by projecting text onto a grid, which LLMs read well, and enables a two-stage pattern: text first, screenshots only when visual reasoning is needed.

Key points

Local execution with no API key or GPU; TypeScript-native with a Python wrapper, on PDF.js and Tesseract.js
Preserves spatial layout on a grid instead of detecting and converting tables, which LLMs read well
Two-stage pattern: text for initial understanding and screenshots only when visual reasoning is needed; output to text or JSON with bounding boxes

Where we saw it

Sam Witteveen 2026-03-26 Sam Witteveen explains that LlamaIndex open-sourced this lighter version of Llama Parse: it runs locally with no GPU or key, is TypeScript on PDF.js and Tesseract.js, keeps layout on a grid and enables the text-first, screenshot-only-when-needed pattern, paying for those calls only when required. LiteParse - The Local Document Parser

Same category

microsoft/markitdown Converts PDFs and office documents to clean Markdown for LLMs Open detail teng-lin/notebooklm-py NotebookLM inside your terminal and your coding agents Open detail