LiteParse (LightParse) is a new open-source document parser from LlamaIndex that runs locally without a GPU or API key. Built natively in TypeScript on PDF.js and Tesseract.js, it supports 50+ file formats and preserves spatial layout by projecting text onto a spatial grid rather than converting tables to markdown — an approach LLMs handle well. It enables a two-stage agent pattern: fast text parsing for most cases, with fallback to screenshot-based multimodal reasoning when needed. The broader context is LlamaIndex's public acknowledgment that the LLM framework era is ending, driven by better agent reasoning, MCP-based tool discovery, and coding agents that can write integration code directly. LlamaIndex is pivoting toward document parsing infrastructure, with LiteParse as the open-source counterpart to their enterprise product LlamaParse.
Sort: