Document intelligence validation
Keep page coordinates beside extracted fields
A reviewer should be able to move from a questionable value to the part of the document that produced it. Preserve that connection through normalisation and correction.
In this article
Store the observation and interpretation separately
For each field, retain the raw text, normalised value, source page and available location information. The raw text might be "1,050.00", while the application stores an amount with currency and minor units. These are related records, not interchangeable representations.
Keep the extraction model and source document version too. If the same file is processed again with a new parser, the result needs a distinct provenance record. Do not overwrite the old result in a way that makes previous review decisions impossible to explain.
Some extraction outputs may lack a precise location or confidence value. Represent that absence explicitly rather than inserting a guessed coordinate or a default confidence that looks authoritative.
Define a field record
A practical record includes the source identity, page index, raw value and interpreted value. Use the coordinate system supplied by the extraction service and retain its units or page dimensions so the viewer can map it correctly.
{
"field": "amountPayable",
"rawText": "AUD 1,050.00",
"value": { "currency": "AUD", "minorUnits": 105000 },
"source": { "documentId": "invoice-fixture-12", "page": 1 },
"extractionVersion": "fixture-parser-v2",
"status": "needsReview"
}The example omits the polygon to keep it readable. In a real implementation, store the returned geometry and its coordinate convention rather than assuming pixel coordinates. The page viewer must apply scaling and rotation consistently.
Preserve provenance during corrections
When a reviewer changes a value, record the corrected value, actor and reason while retaining the extraction observation. This allows later analysis to distinguish model errors from business adjustments.
Do not move the highlight to a different field merely because the corrected value appears elsewhere on the page. If the reviewer selects another source region, record that selection as part of the correction.
Keep derived calculations separate from extracted fields. A calculated total has an arithmetic lineage, while a printed payable amount has a document location. Presenting both as if they were read directly from the page obscures their different origins.
Test the viewer and export paths
Use a rotated scan, a multi-page invoice and a page with repeated totals. Confirm the highlight lands on the intended text at desktop and mobile sizes. Check that zooming does not detach the overlay from the page.
If the result is exported, retain enough provenance to trace it back to the source. A CSV may contain only values, so provide a stable record identifier through which authorised staff can inspect the evidence. The field is operationally useful when someone can verify its meaning without searching an entire document manually.
Primary sources
Microsoft Learn: document layout analysisMicrosoft Learn: invoice extractionReferences checked 11 September 2026.