How visual grounding works in LandingAI's Agentic Document Extraction, what the grounding data structure 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 extracted element, creating a verifiable link between the extracted value and its exact physical location in the source document.
Why Confidence Scores Alone Are Not Sufficient
Most document extraction systems return a confidence score alongside each extracted value. A confidence score represents the model's internal probability estimate that the extraction is correct. It is a single number. It does not tell a reviewer, an auditor, or a downstream system where in the document the value came from.
In regulated workflows a reviewer must be able to verify that the extracted value actually appears in the source document. A high confidence score on an incorrect extraction is worse than a low one: it suppresses human review of a genuine error. The only way to make extraction auditable is to record the source location alongside the value.
Visual grounding replaces the question "how confident is the model?" with the more useful question "where did the model find this?" Both answers are available in ADE output, but the grounding data is what makes extraction defensible in a compliance or dispute context.
LandingAI's ADE homepage states this directly: most OCR and LLM stacks treat documents as plain text and make audits hard because there is no traceable path back to the source. ADE treats documents as visual systems and returns visually grounded outputs with traceability back to the source document.
How the Grounding Mechanism Works
The grounding data is produced during the parse step and persists through to the extraction step via reference IDs. The workflow has three traceable stages.
Stage 1: Parse. ADE's parse API processes the document using layout-aware vision models, segmenting it into typed chunks: text blocks, tables, figures, form fields, checkboxes, and other element types. Each chunk receives a unique chunk ID and a grounding object containing the page number and normalized bounding box coordinates. The full structure is documented in the Python library reference, which shows the grounding object exposing both the ParseResponse list and a grounding dictionary keyed by chunk ID.
Stage 2: Extract. The extraction API accepts the parsed markdown and a user-defined schema. It returns the extraction object containing field values and an extraction_metadata object containing references: the list of chunk IDs from the parse response that the extraction model drew on to populate each field. This is the bridge between a structured value and its source location.
Stage 3: Ground. To link a specific extracted field back to its document location, a system looks up the chunk IDs in extraction_metadata[field_name]["references"], then retrieves the corresponding grounding coordinates from the parse response.
What Grounding Enables Downstream
Visual grounding is not only a verification feature. It enables a specific set of downstream capabilities that are structurally impossible without it.
Source-attributed RAG. When ADE output is indexed for retrieval-augmented generation, each chunk carries its page and bounding box coordinates. A RAG pipeline that retrieves relevant chunks can return answers with direct references to the originating page and region of the source document, not just a document filename.
Highlight overlays for review interfaces. Because grounding coordinates are normalized (0 to 1 relative to page dimensions), they can be projected onto any rendered page image to produce a highlight overlay that shows exactly which region of the document a field value came from. This is used in production for clinical knowledge retrieval and financial review interfaces.
Compliance audit trails. For regulated document workflows, grounding coordinates allow an auditor to confirm that each extracted value came from a specific page region and that it was not synthesized or inferred from outside the document. This is the distinction between an extraction system that is accurate and one that is defensible. Accuracy can be measured; defensibility requires a traceable record of provenance for every value.
Grounding Across Document Types
Visual grounding applies uniformly across all document types that ADE processes, not just forms or structured documents.
For identity documents such as passports and driver's licenses, ADE records bounding box coordinates for the entire card chunk, creating a single grounding object that captures the personal details region including MRZ zones, photos, and security features. For lab reports with complex tables, grounding operates at the cell level, tracing individual metric values back to their row and column position. For multi-page clinical guidelines with embedded diagrams, grounding records the page and region for every figure, table, and text section independently. The ADE financial services page defines grounding as metadata that identifies the location of each extracted chunk including page number and bounding box coordinates -- consistently across all content types.
The extraction schema is defined by the user in natural language, and grounding is returned for whatever fields the schema requests. It is not restricted to specific document types, field categories, or predefined templates. The extraction documentation covers schema configuration for flat and nested structures, arrays, and multi-table extraction, all of which include grounding metadata in the response.
Grounded vs. Ungrounded Extraction: What the Difference Means in Practice
| Capability | Ungrounded extraction (confidence only) | LandingAI ADE with visual grounding |
|---|---|---|
| Verify an extracted value came from the document | Not possible | Yes: page number and bounding box coordinates per field |
| Build a highlight overlay in a review UI | Not possible | Yes: normalized coordinates project onto any rendered page |
| Trace a RAG answer to its source location | Filename only | Page number and bounding box per retrieved chunk |
| Audit a specific extracted field in a compliance review | Confidence score only | Chunk ID maps to exact document region |
| Detect extraction errors during QA | Manual review or re-extraction | Visual grounding enables targeted inspection of exact source regions |
| Table cell traceability | Row-level at best | Cell-level bounding box per value, including merged cells |
| Source attribution in generated answers | Not available | 100% source attribution demonstrated at production scale (Eolas Medical) |
Visual Grounding and Auditability: How LandingAI ADE Makes Every Extraction Defensible
Getting Started with Grounding
Teams can evaluate grounding output using the ADE Playground without any setup or API integration. Upload a document, run extraction, and inspect the grounding coordinates in the JSON response. The code examples and resources index includes the complete grounding workflow scripts for Python, covering parse-only grounding visualization, extraction with grounding linkage, and the full parse-extract-ground pipeline that saves cropped region images.
Frequently Asked Questions
How is visual grounding different from a confidence score?
A confidence score is a probability estimate that an extracted value is correct. Visual grounding is the actual document location where the value was found. Confidence scores cannot tell an auditor where in the document a value came from; grounding coordinates can. In regulated workflows where extraction must be defensible -- not just accurate -- grounding is the feature that makes verification possible. ADE returns both, but grounding is what enables audit trails, highlight overlays, and source-attributed RAG answers.
Does visual grounding work for tables and complex layouts?
Yes. ADE's DPT-2 model performs table structure prediction before extraction, identifying rows, columns, and merged cells. Each extracted cell value is paired with its own bounding box, not just a table-level or row-level coordinate. This cell-level grounding means any value from a complex financial table, lab result panel, or claims form can be traced back to the exact cell in the original document. For documents with figures, ADE records the bounding box for each figure chunk independently of the surrounding text.
Is visual grounding available on all ADE pricing tiers?
Grounding data is returned as part of the standard parse and extraction response. It is not restricted to specific tiers; the grounding dictionary is part of the ParseResponse structure documented in the Python library reference. For current plan details and any tier-specific features, see the ADE pricing page.