Document Pre-Trained Transformers (Parsing Models)

​

Parsing Models Overview

​

Availability

• in the Playground
• when calling the ADE Parse or Parse Jobs APIs
• when using the Python and TypeScript libraries

​

Model Versions and Snapshots

​

Why Model Versioning Matters

1. Use a general model name (like dpt-2 or dpt-2-latest) to always get the newest version. This automatically give you improvements and updates, but parsing results may change when new model versions are released
2. Use a specific snapshot (like dpt-2-20250919) to pin to an exact model version. This keeps parsing results consistent until the snapshot is retired or migrated, but you won’t receive improvements.

​

Understanding Snapshots and -latest

​

DPT-2

• Agentic Table Captioning: DPT-2 can parse large, complex, no-gridline, and merged-cell tables with unprecedented fidelity. Every cell is preserved, aligned, and made accessible—enabling cell-level grounding so you know exactly where values came from.
• Refined Figure Captioning: Logos, seals, and small figures are now identified precisely and concisely, eliminating the noise of verbose descriptions.
• Smarter Layout Detection: Fewer chunks are missed, even in messy scans. DPT-2 can even detect stamps inside tables and process them separately—critical for compliance workflows.
• Expanded Chunk Ontology: Beyond text, tables, and figures, DPT-2 now recognizes attestation (signatures, stamps, seals), ID cards, logos, barcodes, and QR codes—ensuring all document elements are classified consistently. To learn more, go to Chunk Types.

​

dpt-2-20260903

• Output close to earlier snapshots: The snapshot is tuned to match earlier DPT-2 output. Some documents return different text or table formatting.
• No confidence scores: The confidence field in the Parse response is null.

​

When Your Organization Moves

​

dpt-2-20260410

• Improved cell parsing in forms and tables: Text positioned at different locations within a cell is now captured more completely.
• Improved column alignment in complex tables: Cell data now more accurately aligns with its corresponding column headers.

​

dpt-2-20260302

• Table boundary detection: Tables that were previously split into multiple chunks are now correctly identified as a single table.
• Improved large table accuracy: Large tables are now parsed more accurately.
• Special characters returned as Unicode: Characters such as asterisks are now returned as their Unicode characters (for example, *) rather than as spelled-out strings like asterisk.

​

DPT-2 Availability

• ADE Parse
• ADE Parse Jobs

​

Set the Model in the API

curl -X POST 'https://api.va.landing.ai/v1/ade/parse' \
  -H 'Authorization: Bearer YOUR_API_KEY' \
  -F 'document=@document.pdf' \
  -F 'model=dpt-2-latest'

​

Set the Model with the Library

from pathlib import Path
from landingai_ade import LandingAIADE

client = LandingAIADE()

response = client.parse(
    document=Path("/path/to/document.pdf"),
    model="dpt-2-latest"
)

​

Set the Model in the Playground

1. Open a project in the Playground.
2. Click the Parse tab.
3. In the top-right corner of the parse result panel, click Config.
4. From the Model drop-down, select the model you want to use.
5. Click Re-run all files.