Layout - AiQL Handbook

The layout model is a vision detector. It runs on a rasterized page before OCR reads anything. Output is bounding boxes and a region type, not text. It is part of convert on the Middleware API.

Parameters

AiQL’s layout model is a compact object detector. It is not a language model.

Parameter	Value
Kind	Object detector with reading-order prediction
Parameters	33 million
Input	One rasterized page
Output	Bounding boxes, region type, and reading order

Configuration

Setting	How it is configured
Input	Each PDF or image page is rasterized, then detected as a whole page.
Region types	Title, heading, paragraph, table, figure, formula, chart, seal, header, footer, footnote, number, and aside.
Crops	Each box is cropped with a small pad so edges are not cut before OCR.
Order	Layout always runs first. OCR is per region, not a single pass over the full page.
Serving	AiQL selects the checkpoint and thresholds. Customers do not pick a layout model.