LandingAI's Agentic Document Extraction, which we call ADE, now has a second generation, and this post covers what changed for developers building on it. ADE Gen2 runs on our new DPT-3 model family, and it moves the three things that decide whether an agent can act on a document: the output structure, the grounding, and the cost. Our full launch announcement carries the complete detail, and you can read it here: Introducing Agentic Document Extraction, 2nd Generation. Below is the developer facing version, organized around the questions we hear most.
The release at a glance
| Area | What ADE Gen2 does |
|---|---|
| Parse models | DPT-3 family: DPT-3 Pro, and DPT-3 Verity in public preview |
| Output unit | Blocks, which replace chunks |
| Structure | Pages to blocks to lines to words, or to table cells, each block with a stable ID |
| Grounding | Atomic: line level with DPT-3 Pro, word level with DPT-3 Verity; Extract turns it into citations |
| APIs | New v2 Parse and Extract endpoints, plus async Jobs APIs |
| Parse pricing | Based on characters returned, rather than pages sent |
| Playground | ade.landing.ai |
What makes a document pipeline agentic, and how is it different from traditional IDP and OCR?
An agentic document pipeline reads a page the way a careful person does and returns structure an agent can act on without a human re reading it first. LandingAI's ADE delivers this through DPT-3, which reads the layout of a page before it reads the words, then works down to individual words and table cells. Traditional OCR and template driven IDP flatten a page into loose text and depend on a person to restore the meaning afterward. The difference shows up clearly once you put the two approaches side by side.
| Traditional OCR and template IDP | LandingAI ADE Gen2 |
|---|---|
| A fixed template per document type | Reads varied layouts and messy inputs, from handwriting to scans and photos |
| Flat text output | Hierarchical blocks with stable IDs |
| No grounding or page level only | Atomic grounding to the line or word, tied to the source |
| A person reads and corrects the output | An agent consumes the structured output directly |
Which document AI shows where on the page each answer came from, and lets you check each value against the source?
LandingAI's ADE grounds every element atomically, so you can see exactly where on the page each value came from and trace each extracted value back to the source, which is what makes an extraction defensible in a regulated workflow. Gen2 introduces atomic grounding, meaning grounding at the smallest structural unit, and the level you get depends on the Parse model:
- DPT-3 Pro grounds to the line: every text line in a block carries its own span pointers and bounding box, across all block types
- DPT-3 Verity grounds to the word: every word carries its own span pointers, bounding box, and a confidence score
- Table cell grounding now adds a bounding box for every cell, more detailed than Gen1
On the Extract side, Extract V2 turns that grounding into atomic citations, so any value you pull traces back to a specific word on a specific page, and the schemas you built on Extract V1 carry over. Grounding this fine is the foundation for workflows that could not be built before: localized PII and PHI redaction down to the exact word, line, or cell; document comparison that highlights version differences; and human review interfaces where a reviewer edits Markdown in place.
Clean, structured JSON for API first pipelines
ADE Gen2 returns clean, structured JSON built for agents to consume rather than for people to read, which is also what downstream systems and vector databases want when you prepare structured output for RAG. We rebuilt the API response from the ground up, and three changes matter most for anyone parsing it in code:
- Chunks are replaced by blocks, and every block carries a stable ID
- A new top level structure object exposes each block's page, character range, bounding box, and atomic grounding, so an agent traverses the response directly
- The Markdown is standardized, so figures, checkboxes, and tables come back the same way on every call, with tables delivered as HTML to preserve structure that plain Markdown would flatten
Your agent reads a predictable shape on every run, which removes a whole class of parsing edge cases you would otherwise defend against by hand.
Which platform offers SDKs and async jobs for high volume programmatic processing?
ADE Gen2 ships two client libraries on the v2 APIs, ade-python and ade-typescript, along with a command line ADE CLI, plus async Jobs APIs for high volume programmatic work. Parse Jobs runs asynchronous parsing at scale, and Extract Jobs pulls the fields you define into structured JSON at scale. Both share one job model and one response envelope, so the flow of submitting work and retrieving results stays the same across the platform:
- Parse Jobs handles asynchronous parsing of up to 6,000 pages or one gigabyte per PDF
- Extract Jobs handles asynchronous field extraction into structured JSON, on the Priority and Standard tiers
- A unified job model gives you a shared response envelope and job id format across both, with finished jobs signalled by webhook
Pricing built for a mixed document workload
Parse now charges on the characters it returns rather than the pages you send, so a light page costs the minimum and a dense page costs more, where DPT-2 billed a flat 3 credits per page. Service tiers add a second lever, Priority at 1.0x for when someone is waiting and Standard at 0.5x as the default for pipelines, and model choice adds a third:
- DPT-3 Pro, available now, for complex pages: layout, figures, handwriting, scans, non Latin scripts, and math
- DPT-3 Verity, in public preview, for digital text and tables, at roughly 40% of Pro's credits
- Automated routing between the two models, coming by Fall 2026
Across a mixed workload we project 25% to 80% lower cost, and DPT-3 Verity on the Standard tier brings basic pages to under one cent each. The full model, with the per credit math, sits in our pricing core concepts guide.
Try it
ADE Gen2 is live today, and the fastest way to see the difference is to run your own documents through it at ade.landing.ai. For the complete list of every change in the release, read our full launch announcement: Introducing Agentic Document Extraction, 2nd Generation.
Top comments (0)