Build a semantic model by chatting
Turn raw tables into a curated model through conversation.
Raw warehouse tables answer questions only if you already know how they join. A semantic model captures that knowledge once — what the entities are, how they relate, and which curated views matter — so everyone (and the agent) gets clean answers without re-deriving the joins each time. Databasin One can build one with you, just by chatting.
Why a semantic model helps
- Cleaner answers. The agent reasons over well-defined entities and joins instead of guessing how a dozen raw tables fit together — fewer wrong joins, fewer accidental cartesian products.
- Reusable definitions. "Customer," "active subscription," "net revenue" mean the same thing every time, for every question.
- Curated starting points. Gold views package the joins and filters a given audience needs, so common questions are one step away.
Native connectors ship with a semantic model already built — see How connectors work. For everything else, you can grow one conversationally.
How the conversation goes
Building a model is a back-and-forth, not a form:
- You state the questions you want answered. This is the foundation — "what's revenue by region and month?", "which customers churned last quarter?". The agent will ask for them if you jump ahead, because the model is shaped around the questions.
- The agent reads your schema. It introspects the real tables and columns first, so it only models data that actually exists.
- It proposes a model. Entities, the relationships (joins) between them, and persona-driven Gold views — a first draft you can react to.
- You refine in plain language. "Make those left joins." "Add the orders dimension." "Split the finance and ops views." The agent applies each change and shows you the result.
What it builds
The output is a DSM v1 — Databasin's semantic-model format: entities, the joins that connect them, and Gold views aimed at the personas you name. It's built from the same modeling rules the standalone Semantic Model builder uses, so a model you shape in chat follows the same conventions as one you'd build on the canvas.
If you tell the agent who the Gold views are for — "finance leadership," "support ops" — it tailors the curated views to those personas instead of producing one generic catch-all.
Tips for a good model
- Bring real questions, not table names. The questions drive the entities and views; the agent finds the tables.
- Only what exists gets modeled. If a column isn't in the warehouse, it won't be in the model — add the data first.
- Iterate. The first proposal is a draft. The refine step is where it becomes yours.