Ontologies are back, but for the second reason not the first

One of the most watched talks from a recent AI engineering conference was a professor reintroducing ontologies to a room full of people building agents. An ontology, plainly, is a written down description of the things in a domain and how they relate. Classes, properties, relationships. The shorter version offered on stage was data as graphs. The argument is that probabilistic models need what the talk called logical guardrails, and a graph of rules is one.

Two different arguments are riding in that coat, though, and they are not equally strong.

The first is retrieval. Structure your knowledge as a graph and the agent finds the right thing. Maybe. The vault’s own retrieval notes are blunt here: lexical search first, embeddings second, hybrid fusion third, reranking fourth, and graph retrieval only after measured failures. Graph approaches earn their keep on specific shapes of question, entity and relationship queries, corpus level summaries, contradictions, multi hop synthesis. Not as the default layer under everything.

The second argument is verification, and that one I buy. One slide put it as a rule a machine enforces, set against language that is slippery. That is not retrieval at all. That is a checker sitting after the output, deciding whether what came back is even permitted by the domain. A bounded set of rules around an unbounded loop, as the talk framed it.

The honest objection is the one that killed the semantic web the first time. Somebody has to maintain the thing, and nobody wanted to. A developer quoted in the writeup suggests agents could now maintain the ontology as part of their own operation, updating definitions when they hit edge cases, and grants that it is still hard.

So I would not rebuild your search stack around a knowledge graph this week. I would think hard about whether the part of your loop that checks the answer is a model grading a model, which is the usual arrangement, or an actual rule that fails loudly when it is broken.