The slowest part of standing up a semantic layer is not connecting the sources. It is the modelling: naming every entity, deciding which columns are dimensions and which are measures, choosing sensible aggregations, writing the descriptions that make the model self-explanatory, and drawing the typed relationships that let the engine join across sources. Done properly for a real estate of tables, it is weeks of careful work on an empty page.
SemanticFed is designed to hand the first draft to AI. Given a registered data source and its introspected schema, the assist proposes entities with physical bindings, dimensions with data types, metrics with plausible aggregations, and candidate relationships between them — each with a written description. The output is explicitly a draft.
The review boundary is deliberate and permanent. Proposals arrive as unpublished drafts and publishing stays a human act, because publishing bumps a model version and every downstream query, dashboard, and agent inherits the result. An AI that could publish on its own would be able to silently change what a metric means across a whole company. The assist is designed to make modelling fast, not to make it unattended.
There is a security dimension too. An introspected schema is untrusted input: column names and comments come from systems SemanticFed does not control, and a column comment is data, never an instruction to follow. The design treats source metadata accordingly.
This capability is on the roadmap. The typed model it produces into, and the publishing lifecycle that gates it, are real and working today — the generation step is the part still to be built, and it is sequenced behind the platform's shared multi-model AI runtime rather than a bespoke integration in this product.
Because the AI-assisted generation step is not yet built, this story does not yet have a clickable "Do it yourself" web walkthrough.
Ready to make this your story?
