Benchmarks: Answer 99.16% of DocVQA Without Images in QA: Agentic Document ExtractionRead more

Visual Grounding and Auditability: How LandingAI ADE Makes Every Extraction Defensible

Share On :

How visual grounding works in LandingAI's Agentic Document Extraction, what the grounding data contains, why it matters for RAG pipelines and regulated workflows, and how it differs from confidence scores alone.

What Visual Grounding Means in LandingAI ADE

Visual grounding is the bounding box coordinate set and page reference that LandingAI's Agentic Document Extraction (ADE) attaches to every element it detects.

It creates a verifiable link between an extracted value and its exact physical location in the source document.

What ADE returns for every element:

  • The page it appears on
  • The bounding box that encloses it
  • Grounding down to each visual line of text
  • Grounding down to each individual table cell

Why Confidence Scores Alone Are Not Sufficient

Many document extraction systems return a confidence score with each value. ADE returns a location instead.

What a confidence score is: the model's internal probability estimate that an extraction is correct. A single number.

What it cannot do: tell a reviewer, an auditor, or a downstream system where in the document the value came from.

Why that matters in regulated workflows:

  • A reviewer must verify that a value actually appears in the source document
  • A high score on a wrong extraction is worse than a low one, because it suppresses review of a real error
  • A probability cannot be checked; a location can

Visual grounding replaces "how confident is the model?" with "where did the model find this?"

Most OCR and LLM stacks treat documents as plain text, so there is no traceable path back to the source. ADE treats documents as visual systems and grounds every element it detects.

How the Grounding Mechanism Works

Grounding is produced when ADE parses a document and persists through to the values you extract from it.

1. Parsing records location. ADE detects each element on the page and records the page it appears on and the bounding box that encloses it. Text is grounded per visual line. Tables are grounded per cell.

2. Extraction inherits location. When you pull specific fields using a schema, each extracted value keeps a link back to the part of the document it was read from.

3. Following the link returns a region. Any value in your output resolves to a page number and an area on that page.

Grounding is not a separate lookup you build. It travels with the data.

What Grounding Enables Downstream

Source-attributed RAG. Each retrieved element carries its page and bounding box, so a RAG pipeline returns answers referencing the originating page and region, not just a filename.

Highlight overlays for review interfaces. Coordinates are resolution-independent and project onto any rendering of a page. Because text is grounded per line, a highlight narrows from a paragraph to the single line a value came from.

Compliance audit trails. An auditor can confirm each value came from a specific page region and was not synthesized from outside the document. Accuracy can be measured; defensibility requires provenance for every value.

Grounding Across Document Types

Grounding applies uniformly across every document type ADE processes.

Document typeWhat grounding returns
Lab reports and financial statementsCell-level bounding box and grid position, including merged cells
Multi-page clinical guidelinesPage and region for every figure, table, and text block independently
Identity documents and card layoutsThe card detected and grounded as a unit
Documents with signatures or stampsAttestations detected and grounded as their own element type
Documents with barcodes or QR codesScan codes detected and grounded as their own element type

Figures also carry a classification label, so downstream systems can branch on the kind of visual without re-running detection.

Grounding is returned for whatever fields your extraction schema requests. It is not restricted to specific document types, field categories, or predefined templates. See the extraction schema documentation.

Grounded vs. Ungrounded Extraction

CapabilityUngrounded extractionLandingAI ADE with visual grounding
Verify a value came from the documentNot possiblePage number and bounding box per element
Build a highlight overlay in a review UINot possibleCoordinates project onto any rendering of the page
Highlight a single line rather than a paragraphNot possibleText is grounded per visual line
Trace a RAG answer to its source locationFilename onlyPage number and region per retrieved element
Audit a specific field in a compliance reviewNot possibleEvery extracted value resolves to an exact page region
Link an extraction back to the parse that produced itNot availableCarried in the response
Detect extraction errors during QAManual review or re-extractionTargeted inspection of the exact source region
Table cell traceabilityRow-level at bestCell-level bounding box per value, including merged cells
Source attribution in generated answersNot available100% source attribution at production scale (Eolas Medical)

Getting Started with Grounding

Evaluate grounding output in the ADE Playground with no setup or API integration. Upload a document, run a parse, inspect the grounding in the response.

The code examples and resources index includes grounding workflows for the Python and TypeScript libraries: drawing bounding boxes, linking extraction to grounding, and saving cropped region images.

Frequently Asked Questions

How is visual grounding different from a confidence score?

A confidence score is a probability estimate that a value is correct. Visual grounding is the document location where the value was found.

ADE returns grounding, not confidence scores. Every element carries a page reference and bounding box, so any extracted value can be checked against the source rather than trusted on the strength of a number.

Does visual grounding work for tables and complex layouts?

Yes. ADE identifies table structure, including rows, columns, and merged cells, before extraction.

Every cell carries its own bounding box and grid position, not just a table-level or row-level coordinate. Any value from a financial table, lab result panel, or claims form traces back to the exact cell.

Is visual grounding available on all ADE pricing tiers?

Yes. Grounding is part of the standard parse response on every plan. The pricing page lists Visual Grounding on Explore, Team, and Enterprise. It is not tier-gated.

Does the agentic architecture improve accuracy, or only add traceability?

Both.

ADE breaks parsing into steps and verifies intermediate results before returning output. Single-pass OCR has no error-correction mechanism.

The published result is 99.16% on the DocVQA benchmark, 5,286 of 5,331 questions correct, using only ADE's parsed output with no image access during the question-answering step. Full methodology, all 45 errors, and reproducible code are in the DocVQA benchmark post.

Grounding adds the verifiable source link on top of that accuracy. Regulated workflows require both.