- Published on
LandingAI Releases Agentic Document Extraction Gen2 with DPT-3 Pro and DPT-3 Verity
- Authors
- Name
- DeepHorus
LandingAI has launched Agentic Document Extraction (ADE) Gen2, a major upgrade to its document intelligence platform built around the new DPT-3 model family. This release represents a complete architectural shift from the first generation, moving from treating documents as flat lists of chunks to a hierarchical tree structure. The update emphasizes three core themes: affordability, agent-ready outputs, and atomic grounding, aiming to make document parsing more cost-effective and precise for developers and enterprises.
The Gen2 system splits parsing into two distinct models to optimize cost and capability. DPT-3 Verity is designed for deterministic transcription of digitally created documents, providing bounding boxes and confidence scores for every word, ideal for high-volume text and simple forms. DPT-3 Pro, in contrast, analyzes page layout first to detect complex block types like tables, figures, and signatures, handling scanned pages and non-Latin scripts. LandingAI states that DPT-3 Verity charges roughly 40% of the credits of DPT-3 Pro.
A significant change is the new credit-based pricing model, replacing the previous flat per-page rate. Under DPT-3, cost is calculated using both a page component and an output character component. For example, DPT-3 Pro on the priority tier bills 1 credit per page plus 0.5 credits per 1,000 output characters. The company projects this will lead to 25% to 80% cost reductions on mixed workloads, though actual savings will depend on specific document density.
The Parse v2 response structure is now organized with top-level fields for markdown, metadata, and a hierarchical structure of pages and blocks. Block types include text, tables, figures, and attestation elements, each carrying a semantic ID and a grounding object. Markdown output is standardized, with figures using specific tags and tables delivered as HTML to preserve complex formatting.
The most consequential capability is atomic grounding, which provides a visual reference for every piece of extracted information. DPT-3 Pro attaches an entry per visual line, while DPT-3 Verity does so per word, also attaching a confidence score. This allows teams to route uncertain transcriptions to human review based on the per-word confidence signal, enhancing data reliability.
ADE Gen2 is generally available now, with developers able to start for free in the ADE playground. Enterprises can deploy it across various environments, including US or EU clouds, private VPCs on major providers, inside Snowflake, or on-premises in air-gapped setups. LandingAI plans to introduce automated routing between the two parsing models in fall 2026.
