Snowflake Adds AGENTIC Mode to AI_PARSE_DOCUMENT for Charts and Diagrams
A third parsing mode now interprets charts, diagrams and tables into semantic summaries and structured data, moving document AI deeper into the data warehouse.
Snowflake opened public preview of a third mode for its AI_PARSE_DOCUMENT SQL function on 28 October 2026, alongside the existing OCR and LAYOUT modes. The AGENTIC mode detects visual elements such as charts, diagrams, screenshots and tables, then returns a semantic summary and structured details per element: charts yield series and values, while flowcharts yield nodes and their connections. Independent evaluations published 1–2 November recorded character recall of 0.997 on Japanese-language documents and accurate extraction of flowcharts and trees.
The preview carries stated limits: a maximum of 2,000 images per document, no extract_images support, a best-effort output schema requiring client-side validation, and occasional misclassification of tables with enumeration markers. With Databricks offering a similar function, native document understanding is becoming a standard warehouse capability rather than a niche add-on.