What a semantic view really does
What a Snowflake semantic view covers, what you have to spell out, where it stops, and how to build the same layer on Postgres.
Snowflake World Tour hit Berlin on September 29. I wasn’t there, but colleagues came back with one message: the agentic enterprise runs on a proper data foundation, models become a commodity, and the semantic layer is the glue. Snowflake calls the whole stack the control plane for the agentic enterprise.
In Snowflake, that glue is the semantic view. So what does it really do?
Covers
Semantic view
Leaves out
What it does
- Maps physical tables to business entities.
- Declares the join paths, so the agent doesn’t guess them.
- Defines each metric once, so the agent doesn’t invent the formula.
- Carries hints for the model: descriptions, custom instructions, verified question-SQL pairs.
What you have to spell out
- Primary and unique keys for every table.
- Every relationship between tables.
- Every metric formula, and how it aggregates.
- Descriptions in business language. “Sales table” tells the model nothing.
- Defaults: fiscal year, the date filter when the user gives none.
- 10–20 verified queries to start with, per Snowflake’s own guidance.
Where it stops
- It only sees data that already sits in Snowflake.
- Relationships are key joins, nothing more. A table can’t reference itself, so no employee-manager hierarchy. No circular paths.
- An agent can route between semantic views. It doesn’t join across them.
- Snowflake’s guidance is to start with 5–10 tables. This is a domain model, not an enterprise model.
- Someone has to own it and keep it in sync with the schema.
What it doesn’t cover at all
The semantics of connecting data across sources, and ontologies. It won’t tell you that a customer in one system is the account in another. It has no concept hierarchies and no inference. Snowflake’s own engineers build that as a separate layer: node and edge tables with recursive SQL. In their benchmark, a plain semantic view got 50% of answers right, and 78% with the graph and terminology on top.
A semantic view describes data that is already integrated. It doesn’t integrate it.
None of this needs Snowflake
The same layer works on Postgres.
Build it yourself
- Views with business names as the entities.
- Real primary and foreign keys as the join paths.
COMMENT ONfor descriptions.- One table of metric definitions, one of verified question-SQL pairs.
- A tool that hands all of it to the agent before it writes SQL.
Or take it off the shelf, all Apache 2.0
- Cube Core: metrics, joins and access rules in code, served over SQL, REST and GraphQL.
- MetricFlow from dbt: metrics in code, compiled to Postgres SQL. APIs and governance stay in the paid dbt platform.
- Wren AI: text-to-SQL on its own semantic model, plugs into MCP clients.
The hard part of doing it yourself is turning a metric and a set of dimensions into correct SQL. That’s what the tools are for.