A form reviewer looking at AI-extracted invoice or contract data has always had the same problem: the extracted values live in one place, and the proof they're correct — the actual scanned page — lives in another. Getting confident in an AI extraction meant constantly flipping between a PDF viewer and a form, matching numbers up by eye and hoping nothing shifted between tabs.
The File Renderer's Field Extraction mode closes that gap: it puts the document and the AI's extracted regions in the same view, so a reviewer can see exactly where every value came from and correct it on the spot.
Field Extraction lives on the Document Viewer component's properties panel. Toggle it on for any Document Viewer showing a PDF that's been through an OCR/extraction step, and the viewer switches from a plain document preview into a review surface: the PDF renders as before, but now with colored boxes overlaid on every region the extraction step detected.
Each highlight is stored as a normalized region — a position and size expressed as a fraction of the page (0 to 1), not as fixed pixels. That's what lets a highlight stay locked to the exact word or line it belongs to no matter how far a reviewer zooms in, or what size window they're working in. The highlight is drawn as an overlay on top of the rendered page, scaled to match, so it never drifts out of alignment with the text underneath it.
Either direction gets you to the same place: the value and its source, side by side, with no tab-switching.
The floating card that opens on a region isn't read-only — it's an editable field. If the AI misread a handwritten total or a smudged digit, correct it right there and confirm, and the form's underlying value updates immediately. There's no separate "edit mode" to enter and no need to go find the field elsewhere in the form first.
Every highlight is color-coded by the extraction's own confidence score — high-confidence reads are shown in green, lower-confidence reads in red, so a reviewer's eye is naturally drawn to the values most worth double-checking instead of having to verify every field with equal effort. Line-item rows (like invoice line entries) get a dashed border to distinguish repeating structured data from single-value fields.
The highlight coordinates are stored as an explicit data binding on the Document Viewer, independent of which OCR provider produced them. Whether the extraction ran through Textract, Azure Document Intelligence, or another provider behind the scenes, the review experience — highlights, click-to-locate, inline correction — works exactly the same way.
