
Structured Context Retrieval across the Semantic Enrichment Stack — the architectural pattern for AI on knowledge graphs.
Most knowledge graphs contain records. AI keeps querying records. The value lives in the meaning layer underneath.
The semantic layer is the new infrastructure tier. Gartner has named it critical infrastructure by 2030. McKinsey calls it the foundational tier of agentic AI. Foundation Capital characterizes context graphs as a trillion-dollar opportunity. The category isn't in question. The architecture beneath it is.
This page is the conceptual companion to a five-minute lightning talk at KGC 2026. It names the architectural pattern that makes semantic-layer AI actually work — and the distinction that keeps it honest.
A JSON Schema is not an OWL class. Validation is not inference. The formal-semantics tradition is right to insist on this distinction.
But meaning doesn't live at one tier of formality. It sits on a stack — schema, vocabulary, taxonomy, ontology — each adding formal commitment, each making explicit what the layer below leaves implicit.
Most enterprises live at the bottom of this stack. Most ontologists work at the top. The architectural question is how AI queries across it.
This isn't a hierarchy of correctness. It's a continuum of formality. Domain experts encode meaning through DB foreign keys, JSON validators, and content models — they're doing ontological work in a different notation. Librarians shape taxonomies. Ontologists build formal models that support reasoning. All of it is the meaning layer.
Vector retrieval chunks every tier into embeddings — schema, vocabulary, taxonomy, ontology, all flattened into similarity space, every formal commitment lost.
Even tool-calling at the data tier — which is where most knowledge-graph-aware AI sits today — only queries instances. The meaning above goes unconsulted.
The result is AI that's superhuman in narrow tasks and absurd in obvious ones. Bigger context windows that don't improve retrieval. Agents that hallucinate confidently in regulated domains. Pilots that fail without anyone being able to point to why.
It's not a model problem. It's a substrate problem. The fix is upstream of the model.
Three moves. Each one earns its place in the architecture.
Runtime Ontology Schema Editing Through Type Alignment
Aligns across sources at the type level. A field in your JSON Schema can map to a SKOS concept can map to an OWL class — each preserving its formal level.
ROSETTA records which tier each alignment lives at, so nothing pretends to be more formal than it is. Patent-pending.
Validated Substrate
The aligned meaning graph lives in CoreModels with SHACL or ShEx validation enforced.
The substrate isn't just any KG — it's a validated one, with provenance preserved end-to-end. This is the OWL→SHACL migration the standards community has been building toward.
Runtime Querying via Model Context Protocol
The agent traverses at runtime — at the schema tier, the taxonomy tier, the ontology tier. Same substrate, different formal depth.
Every answer paths back to source nodes at the appropriate tier. No similarity guessing. No hallucinated structure.
Vocabularies coexist. Alignment lives at type level, not at terminology.
SHACL constraints answered at write time, not retroactively.
Every answer paths through schema, not similarity. Audit becomes possible.
APIs and agents operate on the meaning graph, not on copies of data.
Alignment happens at type level, not by forcing everyone onto a single vocabulary.
The category isn't a vendor invention. Multiple independent voices have arrived at the same conclusion within twelve months. The talk this page accompanies argues about the architecture beneath the consensus — not whether the consensus exists.
Look up
before you make up.
Schema is not Ontology. The meaning layer is the whole stack. Build there.