Bounding box

A bounding box is the set of coordinates, usually four numbers marking the top-left and bottom-right corners, that pinpoints where on a page a specific piece of extracted data was found. It lets an Intelligent Document Processing (IDP) system show exactly which pixels produced a given value, rather than just returning the value on its own. In practice, it is the visual proof behind every extracted field.

How it works

When a document processing system reads a page, it does not just output text: it tracks position too. Each word, line, or table cell detected by OCR or a vision-language model gets a bounding box, typically expressed as (x1, y1, x2, y2) coordinates relative to the page image, or sometimes as a percentage of page width and height so the box scales regardless of resolution.

During extraction, the system links each output field to the bounding box (or boxes) that generated it. If a model reads "Invoice Total: €4,250.00" from a specific spot on page 2, the extracted value "4250.00" carries a reference back to that exact location. Some documents need multiple boxes per field: a multi-line address, for instance, might span three separate boxes stitched together.

Bounding boxes also support downstream steps. A confidence score can be tied to a specific region, so a reviewer knows not just that a field is uncertain, but where to look on the page to verify it. Table extraction relies heavily on nested bounding boxes: one for the table, one for each row, one for each cell. Without this spatial layer, a system can extract text but cannot tell a human where that text came from.

Why it matters for document processing

Bounding boxes turn extraction from a black box into something auditable. A reviewer can click on a flagged field and see the highlighted region on the original page instead of hunting through a multi-page PDF. This matters in KYC onboarding, where a compliance officer needs to confirm that a date of birth was read correctly, and in insurance claims, where an adjuster wants to check the exact line item that produced a payout figure.

Bounding boxes also matter for building trust in straight-through processing. If a system can show its work, low-confidence fields can be routed to a human with the relevant region already highlighted, cutting review time. Without positional data, exception handling becomes a manual search exercise, which slows things down and increases the chance of missed errors. For regulated industries, this traceability is often what makes automated extraction defensible during an audit, and it is one reason document fraud (see our note on Kbis fraud) is easier to catch when reviewers can see the source region, not just the parsed value.

Related terms

• Confidence score
• Table extraction
• Key-value extraction
• OCR (Optical Character Recognition)
• Data extraction

FAQ

Do all IDP systems generate bounding boxes?

No. Some systems return only extracted values with no spatial reference, which makes verification harder. Systems built for regulated workflows, such as KYC onboarding or contract review, generally include bounding boxes by default because auditors and reviewers need to trace each value back to its source location on the page.

Can a bounding box cover more than one word or field?

Yes. A single box can span a phrase, a table cell, or a full paragraph, depending on how the extraction model groups content. For structured data like tables, systems typically use nested boxes: one for the overall table, and smaller ones for each row and cell within it.