A self-improving context engine

Basedash's semantic layer is built from models. A model is a reusable SQL dataset on one data source with semantic metadata attached: column descriptions and synonyms, measures such as monthly recurring revenue, segments such as active subscriptions, a default time column, a row grain, usage guidance, and relationships to other models. Models appear as read-only virtual tables you query by reference name, like models.subscriptions, and every Basedash AI gets their definitions and verification status when it answers a question.

Yes, alongside it. Connect the warehouse where dbt builds your tables and Basedash queries those tables and views in place. Table and column descriptions in the warehouse, including the ones dbt persists there, are read automatically as AI context. Basedash doesn't parse dbt project files, so teams usually build models on top of their dbt marts to add the measures, segments, and relationships agents use, and can manage those models from CI through the REST API.

Define it as a measure on a model. A measure is a named aggregate expression, such as SUM(monthly_amount) for monthly recurring revenue, and a segment is a named filter, such as status = 'active'. Keep the model's SQL at row grain, one row per subscription for example, and set a default time column so the AI can group the metric by day, week, or month. Write these by hand, or describe the metric and let the AI draft the SQL, measures, and segments for you to review.

Yes. Model SQL can clean, join, rename, and reshape raw tables into analysis-ready datasets, and a model can join other models and tables on the same data source. Each model runs in its source's native SQL dialect. To model across SaaS tools like Stripe and HubSpot, sync them into Basedash Warehouse, where each lands in its own schema on one source, then build models on top.

Without models, an AI rebuilds business logic from raw tables on every question, and small differences add up: a missed filter, a wrong join, revenue counted twice. With models, Basedash gives its agents each model's measures, segments, row grain, relationships, synonyms, and verification status. The AI reuses measure expressions in SELECT and segment expressions in WHERE, joins along declared relationships to avoid fanout, and prefers verified models when more than one could answer.

Members can view and edit models for the data sources they can access, and row-level and object-level access rules still filter the model SQL and metadata each person sees. Every change to SQL or semantic metadata creates a version you can review and restore from History. Only organization admins can verify a model, and editing its SQL, columns, measures, segments, or default time column removes verification until an admin reviews the new version.

Models hold executable logic: SQL datasets, measures, and segments that every query can reuse, so the numbers come out the same. Skills hold judgment: plain-language playbooks, such as how finance reports revenue or how to read an experiment, that the AI loads when a request calls for them. Most teams use both, with models for what to calculate and skills for how to analyze and present it.