OCR - AiQL Handbook

The OCR model is a vision-language reader. It takes the crops produced by layout and turns each region into machine-readable text. Convert on the Middleware API uses it for scans, photographed pages, and native PDFs that still need structure recovered.

Parameters

The OCR reader is a compact vision-language model. Dedicated passes cover text, tables, formulas, charts, and seals. Figure description can use the same multimodal mixture-of-experts family as pipeline.

Parameter	Value
Kind	Vision-language reader
Parameters	0.9 billion
Input	One layout crop
Output	Text, structured tables, formulas, chart values, or a figure description

Configuration

Setting	How it is configured
Routing	Each crop is sent to a task chosen from its layout type, not to one generic prompt.
Text	Body copy, titles, headings, headers, footers, and footnotes use a general reading task.
Tables	A dedicated table pass. Output is structured tables, not a flattened paragraph.
Formulas	A dedicated formula pass. Output is mathematical notation.
Charts	A dedicated chart pass. Output is the chart’s structure and values.
Seals	A dedicated seal pass for stamps and chops.
Figures	Images use a separate vision-language description pass.
Serving	AiQL selects the checkpoint and timeouts. Customers do not pick an OCR model.