A concept becomes useful in two directions: it connects to other concepts through relationships, and it connects to real data through bindings.

Relationships

A relationship is a typed, explicit edge from one concept to another — “Customer places Order”. It exists because someone declared it and a certifier blessed it. There is no reasoner and nothing is ever inferred: an edge you did not declare does not exist, however obviously two concepts seem related. This is meaning, not plumbing. The column pair two datasets actually join on lives on the semantic model’s own relationship, not here — a concept relationship says two business things are connected; a model relationship says how their rows match. You edit a concept’s own outgoing edges from its edit page. An edge whose other end you cannot see does not appear at all, on either end.

Bindings

A binding says where a concept lives: one semantic model that realizes it. A concept can carry any number of bindings, because the same business idea often lives in several datasets at once — a CRM table, an event stream, a history table — which is exactly the constraint a single semantic model cannot express. A concept’s page lists its bindings under Realized by. If nothing binds a concept yet, it is vocabulary with no data behind it — agreed wording that nothing can currently be asked of.

Filtering by table visibility

A binding names a model, and a model names a real table — so a binding is filtered the same way any table reference is. If you cannot see the table a binding points to, that binding is dropped from what you’re shown entirely, never returned with its details blanked out. The concept itself is not affected: a concept name is business vocabulary, not a table name, so hiding the whole concept because one binding is invisible would make the vocabulary look different depending on who’s asking, for no real security benefit. When at least one binding was withheld this way, the page says so in a single sentence with no count — the number itself would give away how much you can’t see.

Adding and removing bindings

Bindings are edited as part of the concept’s whole document. Picking the concept-side property is a closed list — you can only pick a property this concept actually defines, since the server refuses a property map naming one that isn’t. Picking the model-side element is open: you can type a name the model doesn’t currently define, and the server keeps the mapping with a warning rather than rejecting it, since a restricted author may be binding to elements they cannot fully see.
Editing a concept replaces its whole document, bindings included. If any bindings or relationships were withheld from what you were shown (see above), the edit form refuses to save rather than silently deleting the parts you couldn’t see — see Concepts and the ontology.

How agents walk the graph

An agent that reads a word out of a question — “churn” — needs to find out what surrounds it before it can ask anything useful. Rather than fetching one declared edge per round trip, agents can request a bounded neighbourhood around one concept: every concept reachable within a hop limit, the declared edges between them, and which tables they resolve to.
  • The hop limit is capped by the server, not by the caller — up to three hops, regardless of what’s requested. No request can turn this into an unbounded graph walk.
  • Only edges that were actually declared are returned. Nothing is composed into a new fact — if A relates to B and B relates to C, walking from A returns those two edges and says how far C sits, but never invents a direct A-to-C edge.
  • The same visibility rules apply throughout: a concept you cannot see is dropped along with every edge that touches it, and a binding to a table you cannot see is dropped while its concept still appears.
  • The response says whether anything was withheld, and separately, whether the walk stopped because it hit the hop cap rather than because it ran out of graph to walk.
This is the same map a person browses visually on a concept’s page, just reachable in one call instead of one click per edge — which is what makes it practical for an agent to use at all instead of guessing at your vocabulary. See MCP tools.

Ontology concepts

Define business ideas that span more than one dataset, and move them to certified.

MCP tools

Every tool an agent can call, including the ontology-walking ones.