AI review and learning

The AI does two distinct jobs on a document: re-check an extraction on demand, and learn from your corrections so the next document extracts better.

Review by AI

On any review page, Review by AI asks the model to re-extract and cross-check the document:

• Run it before manual editing — it discards manual edits and re-reads the source file.
• A banner at the top of the review page shows the outcome of the latest AI pass, including failures and how long the pass took.
• The AI cross-references the whole document (all pages and attachments of the group), which catches values that a single-page OCR pass misses.

What the AI markers mean

After an AI pass, each field can carry a marker, and a summary chip shows the tally — for example AI review · 3 corrected · 12 confirmed · 2 not in doc:

Marker	Meaning
AI corrected	The AI changed the OCR value. Glance at these first.
AI confirmed	The AI re-read the document and agreed with the OCR.
Not in document	The AI looked for the field and it is genuinely not printed.

This turns review into triage: confirmed fields need no attention, corrected fields deserve a look, and "not in document" tells you to source the value elsewhere instead of hunting the PDF.

Learning from corrections

When you fix a field and save, the correction feeds the correction memory:

• Recurring OCR mistakes (a misread airport code, a systematically wrong consignee) are remembered and applied to future extractions.
• Corrections are per-field and evidence-backed — the goal is that you stop fixing the same mistake twice.

Last updated: 31 October 2026