Lightning‑Fast Open‑Source Local PDF Parser: LiteParse Processes 400‑Page PDFs in 1 Second

Introduction

LiteParse is an open‑source document‑parsing library released by the LlamaIndex team. It is written in Rust, runs entirely locally without cloud dependencies, LLMs, or API keys, and targets fast, lightweight PDF processing.

Core Features

Rust performance base : parses a three‑page PDF in under one second.

Multi‑language bindings : Node.js, Python, Rust, and WebAssembly (WASM) share the same CLI.

Flexible OCR system : built‑in zero‑config Tesseract and optional HTTP OCR servers (EasyOCR, PaddleOCR).

Multiple input formats : PDF, DOCX, XLSX, PPTX, various image types; Office documents are auto‑converted via LibreOffice.

Bounding Box output : each text block includes precise coordinates for downstream AI pipelines.

Agent Skill support : a single command can register LiteParse as a skill for coding agents such as Claude Code, Cursor, and Qoder.

Installation

Choose one of three one‑line commands; all install the same CLI:

# Node.js (recommended)
npm i -g @llamaindex/liteparse
# Python
pip install liteparse
# Rust
cargo install liteparse

After installation, lit --version reports version 2.0.0 (npm registry shows 2.0.4, but the binary reports 2.0.0).

Benchmarks and Usage

Using a real three‑page MiniMax IPO counseling report (Chinese PDF), the author measured:

$ lit parse minimax-ipo-counseling.pdf --no-ocr -o output.txt
[liteparse] extract: 949.4ms (3 pages)
[liteparse] ocr: 0.0ms
[liteparse] project: 3.6ms
[liteparse] total: 953.1ms

The extraction produced 113 lines of text (5 120 bytes), including titles, tables, and company information.

关于 MiniMax Group Inc.
首次公开发行股票并上市辅导备案报告
成立日期 2021 年 6 月 30 日
注册资本 50,000 美元
辅导协议签署时间 2026 年 5 月 29 日

JSON output with bounding boxes ( 47 KB ) is generated in 6 ms when cached:

$ lit parse minimax-ipo-counseling.pdf --format json --no-ocr -o output.json
[liteparse] extract: 5.6ms (3 pages)
[liteparse] total: 6.0ms

When OCR is enabled, the tool intelligently skips OCR for PDFs that already contain extractable text.

$ lit parse minimax-ipo-counseling.pdf --target-pages "1" 
[liteparse] extract: 29.9ms (1 pages)
[liteparse] ocr render: 2.3ms (0 pages)
[liteparse] ocr: 0.0ms
[liteparse] total: 37.8ms

Screenshot generation creates PNGs (1240×1754, 8‑bit RGBA) useful for multimodal LLM workflows:

$ lit screenshot minimax-ipo-counseling.pdf --target-pages "1-3" --dpi 150 -o ./screenshots

Batch parsing recursively scans a directory and processes all PDFs:

$ lit batch-parse ./inputs ./outputs --format text --no-ocr --extension .pdf
[liteparse] found 1 files to process
[liteparse] batch complete: 1 succeeded, 0 failed

Agent Skill Integration

Register LiteParse as an agent skill with a single command:

npx skills add run-llama/llamaparse-agent-skills --skill liteparse

After registration, agents such as Claude Code, Cursor, and Qoder can invoke PDF parsing, screenshot generation, and text extraction directly.

Parse contract PDFs to extract key clauses.

Batch‑generate screenshots for multimodal LLM understanding.

Embed document parsing steps into agent workflows.

OCR Configuration

Built‑in Tesseract works out of the box; language can be specified:

# Chinese
lit parse document.pdf --ocr-language chi_sim
# French
lit parse document.pdf --ocr-language fra
# Disable OCR for pure‑text PDFs
lit parse document.pdf --no-ocr

For higher accuracy, an external OCR server can be used:

# Start PaddleOCR server
cd liteparse/ocr/paddleocr && python server.py
# Use the server
lit parse document.pdf --ocr-server-url http://localhost:8828/ocr

The OCR API expects a POST to /ocr returning { results: [{ text, bbox, confidence }] }.

Pros and Cons

Pros

Extremely fast thanks to Rust; 3‑page PDF under 1 second.

Simple installation via npm, pip, or cargo.

Flexible OCR design with built‑in Tesseract and pluggable external services.

Agent Skill support enhances AI workflow capabilities.

Pure local execution keeps data private.

Cons

Table extraction only reconstructs spatial text; no structured table recognition (requires LlamaParse cloud version for serious table use cases).

Limited handling of multi‑column and complex layouts.

Documentation and CLI parameters for the Skill feature have minor inconsistencies (e.g., --pages vs --target-pages).

Conclusion

LiteParse positions itself as a lightweight, local, and fast document‑parsing foundation, ideal for batch PDF processing, latency‑sensitive pipelines, and privacy‑critical scenarios. It does not aim to solve every parsing challenge, but it excels in the domains it targets and is recommended for RAG preprocessing, agent toolchain construction, and offline document handling.