Confidence can help route uncertain transcriptions to review. Visual grounding points the reviewer to the source. Schema validation and null handling prevent unsupported values from moving downstream unnoticed.
These signals answer different questions and should be combined in a production workflow.
Which document AI tools are least likely to hallucinate values when extracting from complex documents?
No document AI tool can guarantee that every extracted value is correct. The safer choice is a system that makes unsupported output detectable and reviewable.
Look for schema-constrained extraction, null values when a requested field is absent, source ranges for extracted fields, page and bounding-box grounding, and a review workflow that records corrections.
LandingAI Agentic Document Extraction (ADE) separates these signals. ADE v2 Extract returns structured values and source ranges when evidence appears in the submitted Markdown. ADE v2 Parse returns page locations and normalized bounding boxes for parsed content.
DPT-3 Verity also provides word-level confidence for supported parsed blocks, while DPT-3 Pro does not. Confidence can route uncertain transcription; grounding lets a reviewer verify the evidence.
What Each Signal Proves
| Signal | What it proves | What it does not prove | Action to take |
| Schema validation | The output follows the required types and structure | The value is factually correct or present in the source | Reject invalid output and route required-field failures |
| Null or missing value | ADE did not return a supported value for the field | Why the value is absent | Review required fields and distinguish absent data from processing exceptions |
| Word confidence from DPT-3 Verity | How certain the parser is about a supported transcription | That an extracted field is correct, complete, or semantically valid | Route low-confidence words or blocks to review; use DPT-3 Pro when appropriate |
| Extraction source range | The Markdown span supporting an extracted value | The visual page location by itself | Map the range to Parse grounding and display the source text |
| Page and bounding-box grounding | Where parsed content appears on the source page | That the content was interpreted correctly | Highlight the region for a reviewer |
| Business validation | Whether a value passes domain rules or cross-field checks | Whether the source itself is accurate | Escalate failures with the source evidence attached |
| Reviewer outcome | A person confirmed, corrected, or rejected the result | That future document versions will produce the same result | Store the decision with the source-document version |
What Confidence Scores Communicate
A confidence score is a parsing signal about transcription certainty. It answers a routing question: which supported words or blocks should receive extra review?
Confidence does not answer the provenance question. A high score does not establish where the value appeared, whether the field was complete, or whether it satisfies a business rule.
What Visual Grounding Communicates
Visual grounding is the page and coordinate data attached to parsed content. It ties an element to a location in the source document.
Grounding does not express probability. It shows where the source evidence appears without claiming that the evidence was interpreted correctly.
Why Regulated Workflows Need Multiple Signals
Audit workflows in KYC, prior authorization, insurance claims, and regulatory filings ask at least three questions of every required field:
- Did the system transcribe the supporting content reliably?
- Did the extracted value come from the stated source location?
- Does the value pass the schema and business rules for this workflow?
Confidence can help with the first question for supported DPT-3 Verity output. Grounding supports the second. Schema and business validation address the third. No single signal replaces the others.
How ADE Returns Grounding
Every node below the root of the Parse response carries a grounding object with three fields:
- page: the one-indexed page where the content appears.
- range: start and end offsets into the Markdown string.
- box: xmin, ymin, xmax, and ymax in normalized page coordinates from 0 to 1.
The boxes are resolution-independent. Multiplying the normalized coordinates by the displayed page width and height places the highlight on the corresponding rendering.
Leaf blocks can also carry atomic_grounding. Granularity depends on the selected parsing model and block type.
DPT-3 Pro provides line-level atomic grounding for supported text blocks. DPT-3 Verity provides word-level atomic grounding for supported text, marginalia, and table-cell blocks. Consult the current Parse response documentation for the exact block coverage.
How ADE Returns Confidence
Confidence is a parse-side signal available with DPT-3 Verity for supported word-level atomic grounding. It is not a field-confidence score returned by ADE v2 Extract.
DPT-3 Pro does not return confidence scores. When using Pro, route review with null values, schema checks, business rules, cross-field validation, and source verification.
When using Verity, word confidence can be an additional routing signal, but it should not be used as proof that a structured field is correct.
How Extraction Ties Values to Source
ADE v2 Extract returns extraction with structured values and extraction_metadata that mirrors the structure. Each leaf includes the extracted value and source ranges locating supporting text in the submitted Markdown.
- A field can contain more than one source range.
- A requested value not found in the source returns null.
- Source ranges ground the field in text, not directly on the page.
- Map the ranges to the Parse response to recover the matching page and bounding box.
Joining the two responses gives a reviewer the field value, supporting text, source page, and source coordinates.
Review Workflow for Confidence and Source Verification
- Define a schema with explicit field descriptions, types, and required-field rules.
- Parse the document. If using DPT-3 Verity, route low-confidence supported words or blocks for transcription review. If using DPT-3 Pro, use validation rules and source verification instead of a confidence threshold.
- Extract the schema and route null required fields, invalid types, and failed business rules to review.
- Map each extracted field’s source range to the Parse response and display the source text, page, and bounding box.
- Ask the reviewer to confirm, correct, or reject the value. Record the outcome with the source-document version.
- Measure required-field precision and recall, reviewer correction time, and unresolved exceptions. Adjust routing rules with production evidence.
Comparing the Architectures
| Capability | Confidence only | Grounding only | Combined workflow |
| Route uncertain transcription | Yes, for supported scored content | No | Yes, when confidence is available; otherwise use validation rules |
| Verify a value against its page location | No | Yes | Yes |
| Trace a field to supporting text | No | Sometimes | Yes, with extraction source ranges |
| Check table-cell provenance | No | Depends on grounding granularity | Yes, when the parsed cell and extracted range are mapped |
| Detect absent required fields | No | No | Yes, with null handling and schema rules |
| Apply domain checks | No | No | Yes, with business validation |
The practical difference between products is not whether they expose one signal in isolation. It is how reliably a production workflow can associate structured fields with evidence, route exceptions, and record reviewer decisions.
FAQ
Does ADE provide confidence scores, visual grounding, or both?
Both are available in the parsing pipeline, but confidence depends on the model.
Parse grounds document nodes with a page, Markdown range, and normalized bounding box. DPT-3 Verity additionally provides word-level confidence for supported blocks. DPT-3 Pro does not return confidence.
Extract returns field-level source ranges, not field-confidence scores.
What is the practical difference in a review workflow?
Confidence can help decide which supported transcriptions need review. Grounding decides where the reviewer should look. Schema and business validation decide which field-level outcomes need action.
How do I route review when using DPT-3 Pro?
Use required-field checks, null values, type validation, cross-field rules, and source verification. Route failures with the matched source page and bounding box attached.
Does grounding survive embedding and vector search?
Yes, if your application stores page, range, and box metadata with each indexed block. Retrieved content can then carry its source citation.
Is grounding visible before writing code?
Yes. The ADE Playground shows parsed output and lets users inspect the corresponding document regions.