Roundup
Your semantic model now answers in your platform's SQL
The semantic query compiles a question in the Domain platform's dialect, including level, time-series and dimension-rule metrics, on DuckDB, MotherDuck, Microsoft Fabric, Databricks and Snowflake.
A question about revenue used to come back as SQL built for DuckDB. This week it comes back in the dialect of the platform the Domain runs on: DuckDB, MotherDuck, Microsoft Fabric, Databricks or Snowflake.
What got easier
Ask the semantic model and get SQL your platform can run. The semantic query now compiles level, time-series and dimension-rule metrics in the Domain platform's dialect, on DuckDB, MotherDuck, Microsoft Fabric, Databricks and Snowflake. It also refuses a query that references a relation the semantic model does not declare. Run mode executes on DuckDB only; on any other platform, a run call is refused up front and points you to SQL mode for the compiled query.
Try it: in Studio, open an Intent and ask for a governed metric, then read the SQL it compiles.
More things that make your week easier
Your agent can find a curated skill and offer to install it. The agent can search the Vibedata Skill Library, open an entry and read the skill's own steps. A card offers the install, which goes through the same commit gate as any other change. A skill delivered by a plugin stays with that plugin, so the library adds curated skills without duplicating them.
Snowflake Domains can keep secrets in Snowflake. A native secret store keeps credentials inside Snowflake, and the checks at Open and Reconcile name a missing schema grant or a per-secret read the service principal has and the user does not. New Snowflake Domains default to it; No Secret Store stays for legacy and credential-free bindings.
Try it yourself
Audit a semantic model against its SQL and warehouse. It fixes the drift the evidence settles and reports the rest, and that is the check that makes a compiled query worth trusting.
Where data engineering is heading
A semantic model bounds the agent's search. The post times one internal column-mapping task three ways: 38 minutes unbounded, 21 with the map first, 5.3 when candidates were ranked from the semantic model (source: the post). It explains why the model belongs in front of the connection.
A report is not a data product. A report answers one consumer at one moment. The reusable definition underneath it is what every team and agent can read.
New to Vibedata?
Start in the hosted sandbox, or install from your cloud marketplace.
Get the weekly roundup by email
One email a week: what got easier in Vibedata, a recipe to try yourself, and where data engineering is heading.
Or follow the RSS feed.
