Make the Azure Document Intelligence model configurable (prebuilt-layout in addition to prebuilt-read)

[isi4d]

Description

Description

The remote OCR parser currently calls the Azure Document Intelligence prebuilt-read model. I would like to request an optional setting to select the model, e.g. PAPERLESS_REMOTE_OCR_MODEL, defaulting to prebuilt-read so nothing changes for existing users.

Why

1. Table structure is lost. With prebuilt-read, tabular documents come back as a flat list of cell values, one per line, with no relationship between them. A line-item table from a repair quote arrives as: description, part number, quantity, unit price — each on its own line, in reading order. The text is searchable, but the structure needed for any later extraction (totals, amounts, quantities) is gone. prebuilt-layout returns tables as tables.

2. Layout appears to recognise characters better, not just structure. Microsoft's own changelog states that the Layout model received "improvements to the OCR model for scanned text targeting improvements for single characters, boxed text, and dense text documents". That is relevant for forms and handwritten entries, which is where prebuilt-read currently struggles most in my archive. On a printed quote, prebuilt-read misread a licence plate and several phone numbers, while getting VIN, customer numbers and all monetary amounts correct.

Compatibility

Per Microsoft's model overview, searchable PDF output is supported for both prebuilt-read and prebuilt-layout, so the existing archive PDF generation path should be unaffected.

Cost

prebuilt-layout is priced higher per page than prebuilt-read. Keeping prebuilt-read as the default means no existing user sees a cost change; those who need structure can opt in knowingly.

Alternatives considered

Routing the call through a proxy that rewrites the model, and building a patched image. Both work around a one-line difference and create maintenance burden for something the parser could expose directly.

Other

No response