LandingAI Releases Agentic Document Extraction Gen2 with DPT-3 Pro and DPT-3 Verity
LandingAI released Agentic Document Extraction (ADE) Gen2, rebuilt around the new DPT-3 model family, shifting from flat chunk-based parsing to a hierarchical tree structure Two specialized models: DPT-3 Pro (complex layouts, handwriting, non-Latin scripts, LaTeX) and DPT-3 Verity (deterministic transcription for high-volume digital text at ~40% of Pro's cost) Pricing fundamentally changed from flat 3 credits/page to a hybrid model combining per-page and per-output-character rates, with priority
Analysis
TL;DR
- LandingAI released Agentic Document Extraction (ADE) Gen2, rebuilt around the new DPT-3 model family, shifting from flat chunk-based parsing to a hierarchical tree structure
- Two specialized models: DPT-3 Pro (complex layouts, handwriting, non-Latin scripts, LaTeX) and DPT-3 Verity (deterministic transcription for high-volume digital text at ~40% of Pro's cost)
- Pricing fundamentally changed from flat 3 credits/page to a hybrid model combining per-page and per-output-character rates, with priority and standard service tiers
- Atomic grounding now links every extracted field to specific words or lines with bounding boxes, enabling PII redaction, document diffing, and reviewer UIs
- Gen1 client code is incompatible with Gen2 endpoints, requiring migration; automated routing between Pro and Verity is planned for fall 2026
Why It Matters
This represents a significant architectural shift in document intelligence, moving from naive chunk-based extraction to structured, grounded outputs that are genuinely agentic-ready. The dual-model strategy with transparent pricing gives practitioners real control over cost-quality tradeoffs, while atomic grounding solves a critical pain point around traceability and compliance in enterprise document processing pipelines.
Technical Details
- DPT-3 Model Family: DPT-3 Pro handles complex layouts, scanned pages, handwriting, non-Latin scripts, and LaTeX math by reading page layout before words and detecting block types. DPT-3 Verity provides deterministic transcription for digital documents with per-word bounding boxes and confidence scores, targeting high-volume text, tables, and simple form fields
- Parse v2 Response Structure: Three top-level fields—markdown (in reading order), metadata, and structure (a document→pages→blocks tree). Block types include text, table, table_cell, figure, marginalia, attestation, logo, card, and scan_code, each with semantic IDs (type-index format) and grounding objects
- Atomic Grounding: DPT-3 Pro provides line-level grounding; DPT-3 Verity provides word-level grounding with confidence scores (computed as the lowest per-character score within a word). Table cells carry bounding boxes but Pro leaves cell-level atomic grounding empty
- Pricing Model: Priority tier bills 1 credit/page + 0.5 credits per 1,000 output characters (Pro) or 0.3 credits/page + 0.2 credits per 1,000 characters (Verity). Standard tier halves both rates. Synchronous calls always bill at priority rates
- Markdown Standardization: Figures use
<figure type="CHART">with<description>tags, attestations emit stacked labels like[STAMPED][SIGNED], and tables ship as HTML by default to preserve merged cells
Industry Insight
- The character-based pricing model rewards efficiency in output generation—teams should optimize extraction prompts to return only necessary data rather than bulk transcription, as dense pages may now cost more than under flat per-page pricing
- Atomic grounding with confidence scores creates a natural human-in-the-loop routing mechanism: low-confidence words can be automatically flagged for review, reducing manual QA overhead in compliance-heavy workflows
- The Gen1-to-Gen2 incompatibility means organizations should audit their document pipelines now and plan migration, while the planned automated Pro/Verity routing in fall 2026 will eventually eliminate manual model selection overhead
Disclaimer: The above content is generated by AI and is for reference only.