LiteParse

Local document parser built on PDF.js + Tesseract.js. Zero cloud dependencies.

Binary: lit (installed globally via npm) Docs: https://developers.llamaindex.ai/liteparse/

Quick Reference

# Parse a PDF to text (stdout)
lit parse document.pdf

# Parse to file
lit parse document.pdf -o output.txt

# Parse to JSON (includes bounding boxes)
lit parse document.pdf --format json -o output.json

# Specific pages only
lit parse document.pdf --target-pages "1-5,10,15-20"

# No OCR (faster, text-layer PDFs only)
lit parse document.pdf --no-ocr

# Batch parse a directory
lit batch-parse ./input-dir ./output-dir

# Screenshot pages (for vision model input)
lit screenshot document.pdf -o ./screenshots
lit screenshot document.pdf --target-pages "1,3,5" --dpi 300 -o ./screenshots

Output Formats

OCR Behavior

• OCR is on by default via Tesseract.js (downloads ~10MB English data on first run)
• First run will be slow; subsequent runs use cached data
• --no-ocr for pure text-layer PDFs (faster, no network needed)
• For multi-language: --ocr-language fra+eng

Supported File Types

Works natively: PDF

Requires LibreOffice (brew install --cask libreoffice): .docx, .doc, .xlsx, .xls, .pptx, .ppt, .odt, .csv

Requires ImageMagick (brew install imagemagick): .jpg, .png, .gif, .bmp, .tiff, .webp

Installation Notes

• Installed via npm: npm install -g @llamaindex/liteparse
• Brew formula exists (brew tap run-llama/liteparse) but requires current macOS CLT — use npm as primary install path on this machine
• Binary path: /opt/homebrew/bin/lit

Workflow Tips

• For VA forms, job description PDFs, military docs: lit parse file.pdf -o /tmp/output.txt then read into context
• For scanned PDFs (no text layer): OCR is required; complex layouts may degrade — consider LlamaParse cloud for critical docs
• For vision model workflows: use lit screenshot to generate page images, then pass to image tool or similar
• For batch jobs: use lit batch-parse — it reuses the PDF engine across files for efficiency

Limitations

• Complex tables, multi-column layouts, and scanned government forms may produce imperfect output
• LlamaParse (cloud) handles the hard cases: https://cloud.llamaindex.ai
• Max recommended DPI for screenshots: 300 (higher = slower, larger files)

Reference

See references/output-examples.md for sample JSON/text output structure.