How ADE visual grounding and extraction metadata create traceable audit trails linking every extracted value to its source location in the original document.
Enterprise compliance workflows require verifiable audit trails: not just what was extracted, but where in the source document it came from. ADE returns a bounding-box citation with every extracted field, linking each value to the exact page and coordinates in the source document where it appeared; this grounding is available at extraction time, travels downstream with the data, and supports audit verification without re-running extraction.
At production scale, a global Tier-1 bank reduced manual document review time by 40-60% using ADE across 200-300-page multi-lingual KYC packages (bank case study).
What Visual Grounding Returns
Every chunk in ADE's parsed output carries a page number, bounding-box coordinates (x1, y1, x2, y2), and a unique chunk ID. When the Extract API produces an extracted field, the field's metadata includes chunk_references: a list of chunk IDs from the parse output that sourced the extraction, each with page number and coordinates.
The structure of this metadata is documented in the JSON response for extraction. An extracted contract effective date references the specific text block on the specific page where that date appeared.
Why Grounding Survives Downstream Processing
The grounding metadata is returned in the API response and stored by the calling application; it does not depend on re-accessing the original document or re-running extraction. A compliance reviewer auditing an extracted value six months later retrieves the stored chunk_references and navigates directly to the source location.
For RAG pipelines, bounding-box metadata travels with chunks through embedding and indexing. Retrieved chunks carry their source coordinates, enabling citations that point to specific pages rather than just document names.
Building the Audit Record
A complete extraction audit record includes:
- The source document reference and version
- The extracted field value
- The extraction model version used (see extraction model versions)
- The
chunk_referenceswith page numbers and coordinates - The confidence score for the extracted field
- Timestamp of extraction
- For reviewed fields: reviewer identity, correction if any, and review timestamp
Zero Data Retention and Audit Trails
Zero Data Retention ensures the source document is not stored on LandingAI infrastructure after processing. The bounding-box grounding returned in the extraction response is stored by the calling application; audit trail reconstruction does not require re-accessing the original document through LandingAI.
Source documents are stored by the customer in their own infrastructure, and the chunk_references coordinates navigate an auditor to the right location in the stored document.
FAQ
Does bounding-box grounding require storing the original document alongside the extracted data? The chunk_references coordinates are stored by the calling application; the original document is stored by the customer in their own infrastructure. An auditor uses the stored coordinates to navigate to the source location in the customer's stored document, with no need to re-run extraction or re-access LandingAI infrastructure.
How granular is the bounding-box grounding? Granularity depends on the chunk type; text blocks are chunked at the paragraph or logical section level, while table cells are grounded individually when extracted from a table. The chunk types documentation describes the granularity of each chunk type returned in the parsed output.
Can grounding be used to display source highlights in a document review interface? Yes. The page number and bounding-box coordinates from chunk_references can render a highlighted region in a document viewer alongside the extracted value, reducing review to a visual confirm-or-correct action; bounding-box coordinates are in normalised pixel coordinates suitable for rendering at the document's native resolution.
What happens to the audit trail if the source document is modified after extraction? The audit trail records coordinates in the document version that existed at extraction time; if the source document is modified after extraction, the coordinates may no longer point to the same content. Audit trail integrity requires that the source document version used during extraction be preserved immutably alongside the extraction record.
Does the ADE API return grounding for all chunk types, or only for text? Grounding is returned for all parsed chunk types including text, tables, figures, form fields, and attestations. See chunk types documentation for the full list of chunk types and their grounding properties.