Sigma runs a 2,883-model semantic layer as code on dbt and Dagster
Sigma Computing runs its own semantic layer as versioned code inside a dbt project, orchestrated by Dagster across 2,883 dbt models and 6 data regions. Matt Senick, the company's lead data platform engineer, described the setup in a dbt Summit 2026 session, according to a dbt Labs blog post.
- Sigma Computing runs its own semantic layer as versioned code in a dbt project, orchestrated by Dagster.
- 63 Dagster jobs build and deploy across 2,883 dbt models and 6 data regions.
- The dbt packages dbt-sigma-data-model, dbt-semantic-view and dbt-cortex-agent compile definitions into Sigma data models, Snowflake semantic views and Cortex agents.
- Senick previewed Open Semantic Interchange (Ossie), meant to allow conversion between vendor formats.
Why a semantic layer matters for AI
Senick said the semantic layer has three jobs in the age of LLMs. The first is to give AI applications guardrails, such as PII restrictions and SQL formatting rules, built directly into model-level instructions. The second is to make sure every AI surface in the company uses the same definitions for a given metric, so they never disagree with each other. The third is to give agents enough model context, synonyms, and verified queries to move beyond generic text-to-SQL guessing.
Borrowing a framing from Anthropic's published research on the topic, Senick argued that once a question maps cleanly to a defined metric, an agent should be able to call a function and get back the one number every other surface in the company already agrees on.
Semantic objects defined in SQL and Jinja
Inside Sigma's own dbt project, semantic objects are defined and version-controlled directly in SQL and Jinja. They are mapped to specific warehouse tables, with explicit column-level control over what is exposed to business users. Sigma sells the same kind of semantic layer to its customers, according to the post.
The company splits its semantic models by business function, including sales, marketing, customer success, and product. It relies on a small set of dbt packages to compile those definitions:
- dbt-sigma-data-model produces Sigma data models.
- dbt-semantic-view produces Snowflake semantic views.
- dbt-cortex-agent produces Cortex agents.
All of these are handled as ordinary dbt nodes.
Dagster handles the build and deploy loop
Dagster orchestrates the whole system. According to the post, 63 jobs continuously build and deploy across the 2,883 dbt models and 6 data regions. Each run downloads the dbt manifest from S3 and runs a state-aware dbt build. The resulting Sigma data models are then synced back through Sigma's API, and the manifest is re-uploaded for the next run.
Toward portable semantic definitions
Senick also previewed Open Semantic Interchange, an effort nicknamed Ossie. It aims to let teams convert between a Snowflake semantic view, a dbt metric, and a Sigma data model without locking into any single vendor's format.
For data teams, the session offers a concrete example of treating semantic definitions as reviewed, versioned code, with one set of metric definitions feeding BI models, warehouse views, and AI agents. Recordings of the keynotes and breakout sessions from dbt Summit 2026 are available on demand, according to dbt Labs.
dbt Labs: How Sigma runs a 2,883-model semantic layer on dbt and Dagster
