Snowflake explains how it built an internal semantic layer to give AI agents and BI tools a consistent, governed understanding of business data. Semantic views sit between raw tables and consumers, translating physical schemas into standardized facts, dimensions, and metrics, improving accuracy, speed, and cost for AI agents. The post covers internal usage stats (400+ users, 5,400+ queries via a product data science agent in one month), best practices around versioning and testing semantic views with dbt, prioritizing performance with materializations and Dynamic Tables, curating and routing queries with Cortex Agents and Cortex Sense, and interoperability with the open Apache Ossie semantic standard.
Questions this post answers
How much does adding a semantic layer improve text-to-SQL accuracy for AI agents?
Adding semantic context to a data model increased text-to-SQL accuracy from about 20% to more than 90% across 40 business questions tested against the TPC-DS benchmark, according to AtScale's benchmark testing. This illustrates why enterprises building AI agents on top of raw tables often see poor query accuracy without a semantic translation layer. Anyone evaluating semantic layers for AI accuracy can track real-world benchmarks like this on daily.dev.
What is a semantic view in Snowflake and how does it help AI agents query data?
A semantic view is a top-level Snowflake object that sits between raw physical tables and downstream consumers like dashboards or AI agents, translating governed business language into database schemas made of facts, dimensions, and metrics. This lets agents skip time-consuming table discovery and sampling, execute SQL directly, and use materialized preaggregated data for lower token cost and latency. Teams designing agent-ready data architectures can follow semantic layer patterns like this on daily.dev.