LandingAI's DPT-3 Shifts Document Extraction From

LandingAI shipped a second-generation document-extraction system this week. It is the shape of the output.

According to marktechpost.com, the redesign, announced September 9, replaces chunk-based document handling with a hierarchy of documents, pages and blocks, and splits the product into two variants: DPT-3 Pro, which grounds every extracted answer to a line of the source, and DPT-3 Verity, which grounds answers to individual words and attaches a confidence score to each one. LandingAI's own announcement puts it bluntly: "Citations tie every extraction back to its exact place in the source," and in Gen2 the landing.ai team describes DPT-3 Pro's citations as atomic, grounded at the smallest structural units of a document rather than at the level of a whole page or paragraph.

That is a real architectural commitment, and it tells you where this corner of the market is heading. Document AI spent the last few years obsessing over producing clean text. The Gen2 design assumes clean text is table stakes and competes on provenance instead: who can show you not just the answer, but the receipt.

Why Chunks Lost

The chunk was always a compromise. Retrieval systems inherited it from search indexing, where slicing a document into fixed-size passages worked well enough for ranking. But extraction is a different job. When you ask a system to pull the counterparty name from a lease, the liability cap from an insurance policy, or the invoice total from a scanned PDF, you are asking questions whose answers live inside a structure someone deliberately built: sections, tables, line items, signatures.

Chunking flattens that structure. A clause split across a page boundary becomes two fragments with no relationship to each other. A table becomes a smear of text. Downstream models then hallucinate confidence about material that was mangled before the model ever saw it.

The Gen2 hierarchy of documents, pages and blocks, as described by marktechpost.com, restores the structure. That matters for anyone who has tried to build review workflows on top of chunk-based extraction and discovered that the reviewer cannot tell which paragraph an answer came from without opening the original file and playing find-the-sentence.

The official documentation adds implementation-level detail on the DPT-3 parsing models behind this structured-document approach, and it is the right place to look if you want to know what actually runs under the hood rather than what the marketing page implies.

The Two-Variant Bet

The Pro/Verity split is the commercially interesting decision. LandingAI is essentially selling two levels of auditability at two price points. Pro gives you line-level citations: the reviewer clicks, sees the line, verifies. Verity goes further, down to the word, and adds a confidence score per answer.

Word-level grounding is more useful than it sounds in a compliance context. Line-level citation tells you where an answer came from; word-level citation tells you exactly which span the model used, which matters when a line contains three numbers and the extraction picked one. If the model cited the wrong number, a line-level citation looks correct at a glance and wrong on a second look. Word-level makes the error visible in one click.

The confidence score is the riskier half of the proposition. Research on confidence signals for extraction is a live field, and the numbers deserve scrutiny. A study published via ApplyBoard's research page examines embedding similarity as a model-agnostic confidence score for document entity extraction, one of several approaches being tested to see whether model self-assessment can be trusted for triage. A separate benchmark on arXiv tackles calibration directly: the empirical question of whether a score of 0.9 actually corresponds to being right 90 percent of the time. The answer in the broader literature is usually no, at least out of the box. Uncalibrated confidence is worse than no confidence, because it teaches reviewers to trust the wrong fields.

Whether Verity's scores are calibrated, and against what error distribution, is not stated in the available material. The record here is thin, and I will say so plainly: LandingAI has not published a calibration curve or an error breakdown by document type in anything I have seen cited. Teams evaluating the product should ask for exactly that.

The Migration Bill

Now the operational part, which is where vendor announcements get quiet and customer spreadsheets get loud.

Gen1 code will not run against Gen2 endpoints. That is a breaking migration, not an upgrade flag. Any team with production integrations, and document extraction integrations tend to accrete like sediment, will be rewriting calls, re-mapping outputs and re-running validation suites.

Billing changes too. Gen2 counts output characters rather than flat pages. This is a different economic model, and depending on your documents it cuts either way. A pipeline that extracts dense tables from a handful of long contracts may find character billing more expensive than page billing. A pipeline processing thousands of mostly-blank forms may find it cheaper. Flat-page billing had the virtue of predictability; character billing aligns cost with output volume, which is fairer in principle and harder to budget in practice. Teams should run their actual corpus through both models before assuming anything.

Product Hunt's listing for extraction under the Agentic Document Extraction name gives a sense of how the product is being positioned to the builder market, where migration friction and pricing surprises get reviewed in public.

The Broader Pattern

Step back and the Gen2 release fits a recognizable arc. Database vendors spent decades moving from unstructured blobs to structured schemas because structure is what makes data auditable and queryable. Document AI is now retracing that path: first the race was to read anything, next the race is to say where everything came from.

The driver is not technical elegance. It is liability. The moment extracted outputs feed underwriting decisions, legal filings, clinical records or regulatory submissions, a bare answer with no citation is unusable, because no human can certify it. Provenance turns extraction from a text generator into evidence production. Vendors who figure this out first sell into the highest-value workflows; vendors who sell clean text alone are commoditized.

The open question is whether confidence scoring follows the same path to maturity that provenance just did. Citations are cheap to get right: the model either points at the right span or it does not, and a reviewer can check. Calibration is harder, because it requires measuring the model against ground truth across document types, layouts and quality levels, then publishing honest numbers. LandingAI has made the architectural move. Whether it backs the confidence scores with the empirical work that makes them meaningful is the thing to watch, and the thing to ask about, before routing a compliance pipeline through Verity.

The chunk era built systems that could read documents. The evidence era, if Gen2 is a sign of it, will be judged by whether systems can show their work.

Marcus Chen-Ramirez, Senior Technology Correspondent