LandingAI Releases Agentic Document Extraction Gen2 with DPT-3 Pro and DPT-3 Verity
What changed
LandingAI released Agentic Document Extraction Gen2, a complete rebuild of its document processing stack leveraging the DPT-3 model family. Unlike the first generation, which chunked documents, Gen2 restructures data into a hierarchy of document, page, and block. The update introduces two new models: DPT-3 Pro, which aligns extracted data precisely to document lines, and DPT-3 Verity, which grounds outputs to individual words with a confidence score. Additionally, billing for the parsing service now counts output characters instead of flat page counts. Notably, Gen1 code is incompatible with Gen2 endpoints and will not function across generations.
Why builders should care
The move from chunk-based to tree-structured document parsing improves granularity and traceability. For builders, this means more control and precision in how extracted data maps to the original document layout. The line-level grounding of DPT-3 Pro benefits workflows requiring exact positioning, such as financial statements or legal contracts. DPT-3 Verity’s word-level confidence scores enable better quality assurance and conditional logic based on extraction reliability. The switch to character-based billing impacts cost modeling and may require budget adjustments for high-volume users. Incompatibility with Gen1 calls for code migration, which increases integration effort but prevents confusion between versions.
The practical takeaway
Operators using agentic document extraction can expect higher fidelity and explainability out of the box. Those extracting complex documents with detailed formatting will gain tighter alignment between output and source, easing downstream validation and manual review. Confidence scoring helps prioritize or reject uncertain text, improving overall automation accuracy. Budgeting needs reevaluation due to the new billing units. The need to rewrite or update integrations is an operational overhead but also enforces a clean break for deploying improved models without legacy compromises.
What to watch next
Track how builders adopt the new tree-based extraction model and whether they encounter challenges migrating code from Gen1. Watch how pricing changes affect user behavior and volumes. Also, observe if LandingAI expands on word-level confidence features or adds tuning options to balance precision and processing speed. Finally, see if competitors respond with similarly fine-grained document extraction or revert to simpler chunk approaches, which could indicate trade-offs in complexity versus scalability.
AI Quick Briefs Editorial Desk