Instance data vs ontology layer

A mature KG typically has two logical layers that coexist in the same graph but serve different purposes:

By storing them in separate named graphs (or using naming conventions), you gain:

• Independent versioning: evolve the ontology without touching instance data, and vice‑versa.
• Reasoning: use the ontology to infer new instance triples (e.g., from :Alice rdf:type :Employee and :Employee subClassOf :Person infer :Alice rdf:type :Person).
• Reusability: share the same ontology across multiple instance graphs (e.g., in different departments).

A common anti‑pattern is to conflate the two: e.g., storing class hierarchies as application‑specific strings inside instance nodes. This breaks the uniform triple model and prevents generic reasoning tools from working.

Two layers in one KG: instances on the bottom, ontology on the top. An LLM that retrieves both layers can answer questions the instance data alone can't (e.g. is Alice an Agent? — yes, via the subClassOf chain).