LiteParse — PDF Parsing & Visual Extraction

Fast, local PDF parser (LlamaIndex) for text extraction with bounding box coordinates, page screenshots, and visual element detection. Wraps a Node.js CLI (lit) via a Python wrapper.

Installation

uv pip install liteparse
# CLI available as `lit` (requires Node.js/npx)

Current version: 1.2.1. Docs: https://developers.llamaindex.ai/liteparse/

CLI Usage

# Parse PDF to text
lit parse document.pdf

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

# Parse specific pages, no OCR (native PDFs)
lit parse document.pdf --format json --no-ocr --target-pages "1-5,10"

# Screenshot pages at high DPI
lit screenshot document.pdf -o ./screenshots --dpi 200

# Batch parse a directory
lit batch-parse ./pdfs ./outputs --recursive --extension ".pdf"

CLI Flags (lit parse)

Flag	Default	Description
-o, --output	stdout	Output file path
--format	text	text or json
--no-ocr	false	Disable OCR (use for native/digital PDFs)
--ocr-language	en	Language code (is for Icelandic)
--target-pages	all	Page ranges: "1-5,10,15-20"
--dpi	150	Rendering resolution
--num-workers	CPU-1	Parallel OCR workers
--max-pages	10000	Page limit
--password		For encrypted PDFs
-q, --quiet	false	Suppress progress

Python API

from liteparse import LiteParse

parser = LiteParse()

# Parse — use ocr_enabled=False for native PDFs (much faster)
result = parser.parse(
    "document.pdf",
    ocr_enabled=False,       # Skip OCR for digital PDFs
    ocr_language="is",       # Icelandic OCR when needed
    target_pages="1-10",     # Optional page selection
    dpi=150,                 # Resolution for OCR
)

# Full text
print(result.text)

# Per-page access
for page in result.pages:
    print(f"Page {page.pageNum}: {page.width}x{page.height}")
    print(f"  Text items: {len(page.textItems)}")
    print(f"  Bounding boxes: {len(page.boundingBoxes)}")

Data Model

ParseResult

• .text — full document text
• .pages — list of ParsedPage

ParsedPage

• .pageNum — 1-indexed page number
• .width, .height — page dimensions (PDF points)
• .text — page text
• .textItems — list of TextItem with position data
• .boundingBoxes — list of BoundingBox regions

TextItem — individual text fragment with coordinates

• .text — the text content
• .x, .y — position (top-left origin, PDF points)
• .width, .height — dimensions
• .fontName — font identifier
• .fontSize — font size
• .confidence — OCR confidence (when OCR enabled)

BoundingBox — rectangular region

• .x1, .y1, .x2, .y2 — corner coordinates

Screenshots

result = parser.screenshot(
    "document.pdf",
    output_dir="./screenshots",
    target_pages="1,5,10",
    dpi=200,
    image_format="png",     # or "jpg"
)

for s in result.screenshots:
    print(f"Page {s.page_num}: {s.image_path}")

ScreenshotBatchResult

• .screenshots — list of ScreenshotResult
  • .page_num, .image_path, .image_bytes

Combining with pdfplumber for Visual Detection

liteparse extracts text with coordinates. pdfplumber detects vector graphics (rects, curves, lines) and embedded images. Combine both for full visual element extraction:

import liteparse
import pdfplumber

parser = liteparse.LiteParse()
result = parser.parse("doc.pdf", ocr_enabled=False)

with pdfplumber.open("doc.pdf") as pdf:
    for lp_page in result.pages:
        pp_page = pdf.pages[lp_page.pageNum - 1]

        images = pp_page.images or []      # Embedded raster images
        rects = pp_page.rects or []        # Rectangles
        curves = pp_page.curves or []      # Bezier curves (SVG-like)
        lines = pp_page.lines or []        # Line segments

        # Page classification heuristic
        text_len = len(lp_page.text.strip())
        if len(curves) > 10 and not images:
            page_type = "infographic"      # Vector graphics
        elif images and text_len < 200:
            page_type = "photo_page"
        elif images:
            page_type = "mixed"            # Text + images
        else:
            page_type = "text"

Caveats

• OCR is slow: ~3-5 sec/page. Use ocr_enabled=False for native/digital PDFs.
• Text items are fragments: Words/phrases, not lines. Reconstruct lines by grouping items with similar y coordinates.
• Coordinates: Origin is top-left, units are PDF points (1/72 inch). Page 16 at width=595 is A4 width.
• Icons/emojis: Vector icon fonts render as empty strings or garbage characters in textItems.
• No image extraction: liteparse doesn't extract embedded images as files. Use pdfplumber or pdfimages CLI for that.
• Under the hood: Python wrapper calls npx @llamaindex/liteparse — requires Node.js.