LandingAI Releases Agentic Document Extraction Gen2 with DPT-3 Pro and DPT-3 Verity
LandingAI has shipped Agentic Document Extraction (ADE) Gen2, a rebuild of its doc intelligence stack round a brand new mannequin household known as DPT-3. Gen1 handled a doc as a flat checklist of chunks. Gen2 treats it as a tree, costs it by the characters it returns slightly than by the web page, and grounds each reply again to a particular line or phrase on the web page. LandingAI workforce frames the discharge round three themes: affordability, agent prepared outputs, and atomic grounding.
Is it deployable? Yes. ADE Gen2 is mostly out there now. Developers begin free within the ADE playground. Enterprises can run it in US or EU cloud, in their very own VPC on AWS, Azure or Google Cloud, inside Snowflake, or on premises together with air gapped environments.
Two parsing fashions as a substitute of 1
Gen2 splits parsing into two fashions so the workload picks the value. DPT-3 Verity transcribes digitally created paperwork deterministically and returns a bounding field and a confidence rating for each phrase. It targets excessive quantity textual content, tables and easy kind fields. DPT-3 Pro reads web page structure earlier than phrases, detects block sorts from tables and figures via marginalia and signatures, returns them in studying order, and handles scanned pages, handwriting, non Latin scripts and LaTeX math. LandingAI workforce states DPT-3 Verity costs roughly 40% of the credit DPT-3 Pro costs, and additionally state automated routing between the 2 is deliberate for fall 2026.
The pricing change is the actual story
Under DPT-2, each web page price a flat 3 credit. Under DPT-3, credit consumption is the sum of a web page element and an output character element. On the precedence tier, DPT-3 Pro payments 1 credit score per web page plus 0.5 credit per 1,000 output characters. DPT-3 Verity payments 0.3 credit per web page plus 0.2 credit per 1,000 output characters. The normal tier halves each charges, so a 12 web page Pro parse returning 48,120 characters involves 36.1 credit on precedence and roughly half that on normal. Totals spherical as much as the closest 0.1 credit score, and the response metadata studies each enter to the calculation.
Service tiers are the second lever. Priority is for when an individual or an agent is ready. Standard runs asynchronously at 0.5x worth and fits pipelines that tolerate minutes to hours. Note that synchronous calls at all times invoice at precedence, and the playground itself runs on precedence. LandingAI tasks 25% to 80% price reductions on combined workloads and claims parsing beneath one cent per web page with Verity on normal. Treat these as vendor figures till you benchmark your personal doc combine, because the character element means a dense web page can price greater than it did earlier than.
Blocks, not chunks
The Parse v2 response has three prime degree fields: markdown in studying order, metadata, and construction. The construction is a doc node whose youngsters are pages, whose youngsters are blocks. Block sorts embody textual content, desk, table_cell, determine, marginalia, attestation, brand, card and scan_code. Every block carries a semantic ID within the kind type-index, secure inside a response however not throughout re parses, plus a grounding object with the web page quantity, a spread into the markdown string, and a normalized bounding field.
Markdown output is standardized too. Figures use <determine sort="CHART"> type parts with generated prose remoted inside <description> tags, so transcription isn’t mistaken for mannequin commentary. Attestations emit stacked labels equivalent to [STAMPED][SIGNED], with [ILLEGIBLE_SIGNATURE] and [ILLEGIBLE_TEXT] as fastened literals. Tables ship as HTML by default to protect merged cells.
Atomic grounding and what it unlocks
Atomic grounding is the discharge’s most consequential functionality. Every leaf block carries an atomic_grounding array: one entry per visible line with DPT-3 Pro, one entry per phrase with DPT-3 Verity. Verity attaches a confidence worth from 0 to 1 per phrase, computed because the lowest per character rating in that phrase, giving groups a sign for routing unsure transcriptions to evaluate. Table cells now carry their very own bounding containers, although Pro leaves cell degree atomic grounding empty. Extract V2 attracts citations from this grounding, so an extracted discipline traces to a particular phrase on a particular web page. That makes PII redaction by coordinate, doc diffing and reviewer UIs buildable slightly than approximate.