A semantic model binds exactly one dataset — so “Customer” living in a CRM table, an event stream and a history table is three unrelated models with nothing recording that they’re the same thing. A concept is the object that fixes that: a named business idea with typed properties, explicit relationships to other concepts, and bindings to every semantic model that realizes it. An ontology is your tenant’s set of concepts and the relationships between them. Concepts live on the Ontologies tab of Semantic layer, since they share its entitlement, but they are governed separately from semantic models. Being able to author a metric does not make you able to author what the business calls things, and vice versa — the two are checked as different permissions.

Creating a concept

Create one from the Ontologies tab with New concept. Like a model, this is a full page rather than a dialog, since a concept’s properties, relationships and bindings all grow without limit.

Properties

Each property is a typed attribute of the concept.
A classification label is advisory only. It is stored, exported and displayed, but nothing in the platform’s access control reads it — it does not mask or filter any column on its own. Column masking is a separate grant, configured in Roles & permissions.
An ontology itself (the namespace a concept belongs to) has no certification lifecycle of its own — there is nothing about a namespace to certify. The concept is the thing that gets versioned and blessed.

Managing ontologies

An ontology is created the moment you define a concept under a name that doesn’t exist yet — pick New ontology… in the concept form’s Ontology field. To see every ontology in one place, open Manage ontologies from the Ontologies tab. It lists each ontology with the number of concepts it holds that you may see, and is where you remove one you no longer need. Deleting an ontology is allowed only when it holds no concepts. Removing the namespace is deliberately never a way to delete the certified vocabulary inside it, so an ontology that still holds concepts is refused — you delete or move each concept out first, and then the empty namespace. Managing ontologies is governed with the same permission as concepts, so anyone who can delete a concept can delete an empty ontology.
The concept count shown is what you may see. An ontology can look empty to you while still holding concepts you can’t read — in that case the delete is refused rather than performed, and the page shows the reason.

Certification: draft, proposed, certified, deprecated

A concept moves through the same four states as a semantic model, and the same rules apply: proposing snapshots the current definition for review; certifying names an exact version to bless; editing a proposed concept withdraws it from review; a certified concept under re-review keeps serving its previously certified version until the new one is certified. Deleting a concept is refused, by name, while another concept’s relationship still points at it — you have to remove or redirect that edge first, rather than the platform silently cascading the deletion.

Visibility

Two separate things can be hidden from you when you read a concept, and the page tells you when either happened, without saying how much:
  • The concept itself. If you cannot see a concept at all, it behaves as though it never existed — asking for it by name returns the same not-found response either way.
  • A binding. A binding names a semantic model, which names a real table. If you cannot see that table, the binding is dropped from what you’re shown, but the concept itself still appears — a concept name is business vocabulary, not a table name, so hiding the whole concept over one invisible binding would make the vocabulary unstable depending on who’s looking.
  • A relationship. If a relationship points at a concept you cannot see, that edge is dropped along with it, so it looks as if the concept simply has no such connection.
When something was withheld this way, the concept’s page says so in a plain sentence, with no count — the exact number would leak how much of the graph you can’t see.

Editing a narrowed concept is refused

Editing a concept replaces its whole document — properties, relationships and bindings included — so if bindings or relationships were withheld from what you were shown, saving would silently delete the parts you couldn’t see. The save is refused whenever either flag is set. This is enforced by the service itself, not only by the console form, so an edit made directly through the API or dip is refused the same way rather than quietly dropping the hidden parts. If you hit this, ask an administrator for access to the missing tables or concepts first, then edit the concept with the full document in view.

The graph view

A concept’s page shows a small graph of its immediate neighbours: the concept, its declared relationships, and what they connect to. A single concept with no edges is shown as a card rather than a graph — a graph needs at least two connected nodes to mean anything.
This is a picture drawn from the same ontology graph the console reads for the whole tenant. Explicit relationships only, never inferred — an edge exists because a person declared it and a certifier blessed it.

What agents see

Agents read the ontology through MCP tools rather than the console. They get the vocabulary map — concept names, properties, relationships — and can walk a bounded neighbourhood outward from one concept to find out what surrounds a word before asking a question. They see no table names unless they specifically ask to resolve a concept down to what they can query, and every step is filtered by the same visibility rules that apply here. See MCP tools.

Relationships and bindings

How concepts connect to each other and to the semantic models that realize them.

Export and import

Take a concept or ontology out as an OSI document, and bring one back in.