LandingAI Releases Agentic Document Extraction Gen2 with DPT-3 Pro and DPT-3 Verity

Details

LandingAI has released Agentic Document Extraction Gen2, rebuilding its document intelligence platform around a new DPT-3 model family. The core architectural shift moves from flat chunk lists to a hierarchical block tree (pages → typed blocks: tables, figures, handwriting, signatures), enabling more precise downstream parsing. Two models handle different workloads: DPT-3 Pro tackles scanned pages, handwriting, non-Latin scripts, and LaTeX math; DPT-3 Verity handles digitally-created documents deterministically with a bounding box and confidence score per word. The pricing change is the headline. Gen1 billed 3 credits per page flat. Gen2 bills a page component plus output characters: Pro runs 1 credit/page + 0.5 credits/1,000 chars on priority tier; Verity runs 0.3 credits/page + 0.2 credits/1,000 chars — roughly 40% of Pro's cost. An async standard tier halves both rates. LandingAI claims 25–80% cost reductions on mixed workloads, but dense documents can cost more than before, so benchmarking against your actual document mix before migrating is essential. Word-level grounding — bounding boxes and confidence scores on every extracted token — is a meaningful capability addition for RAG pipelines and compliance workflows that need to cite specific source locations. The structured JSON response (markdown, metadata, block tree) is designed for agentic consumption. ADE Gen2 is generally available now across US/EU cloud, AWS/Azure/GCP VPC deployments, Snowflake, and air-gapped on-premises environments. Automated Pro/Verity routing is planned for fall 2026; developers currently select the model manually.