Reading Order

The Problem

PDF documents don’t store text in reading order. A two-column academic paper has text positioned absolutely on the page — the PDF viewer renders it visually, but the underlying data is a jumble of coordinates.

XY-Cut++ Algorithm

EdgeParse uses an enhanced XY-Cut algorithm to determine reading order:

1. Recursive splitting — divide the page into horizontal and vertical strips
2. Column detection — identify column boundaries from whitespace gaps
3. Block ordering — sort blocks within columns top-to-bottom, left-to-right
4. Cross-column flow — handle elements that span multiple columns

How It Works

Page Layout          XY-Cut Analysis       Reading Order
┌────┬────┐         ┌────┬────┐          1. Title (span)
│ Title    │         │ █████████│          2. Col 1, Para 1
├────┼────┤  ──▶    ├──▶─┼──▶─┤   ──▶   3. Col 1, Para 2
│Par1│Par3│         │ 1  │ 3  │          4. Col 2, Para 3
│Par2│Par4│         │ 2  │ 4  │          5. Col 2, Para 4
└────┴────┘         └────┴────┘          6. Footer (span)

Benchmark

EdgeParse achieves a NID score of 0.889 on 200 diverse documents — the highest reading-order accuracy in the current benchmark snapshot.

Tool	NID Score
EdgeParse	0.889
OpenDataLoader	0.873
Docling	0.867
PyMuPDF4LLM	0.852