Snowflake Document AI is powered by Snowflake’s proprietary foundation model, Arctic-TILT (Text Image Layout Transformer). Arctic-TILT is a multimodal architecture optimized specifically for visually rich document understanding, information extraction, and visual document question answering.
Unlike standard text-only large language models that process linearized token streams, Arctic-TILT simultaneously processes three distinct modal signals within a document:
Textual Content: The raw characters and words identified within the document.
Visual Imagery: Visual cues such as logos, check boxes, stamps, and signatures.
Spatial Layout: Precise 2D geometric coordinates, bounding boxes, and tabular relationships (e.g., column headers aligning with invoice row items).
By unifying text, image, and spatial layout modalities within a single architecture, Arctic-TILT extracts data points from complex unstructured and semi-structured documents (e.g., PDFs, scanned invoices, contracts, receipts) with high precision. General-purpose models such as mistral-large2, deepseek-rl, or jamba-instruct lack this native 2D layout transformer pipeline and are deployed in Cortex for general reasoning and completions rather than Document AI extraction.
References & Study Guide Topics:
Snowflake Document Processing: Arctic-TILT Multimodal Architecture
Document Understanding: Spatial, Visual, and Textual Extraction in Document AI
===============
Submit