09 · Why ontology, not retrieval
Retrieval finds what sounds right. Ontology knows what is true.
A language model is the ultimate form of semantic ontology: it knows the word for a tree, and that does not let it rebuild one. Retrieval and knowledge graphs give it more words. An ontology gives it the thing itself, which is why it is the seed of safe control and automation for agents.
RAG and vector searchKnowledge graphOntology
What it holdsPassages that sound relevant to the question.
Things, and the links between them.
What exists, what it is made of, why, how it relates, what must be true, and where an agent may act.
When it is wrongNothing notices. Similar is not the same as true.
It goes stale unless someone maintains it.
A change that breaks a claim is caught, and the map moves with the work.
Can you rebuild from it?No.
Partly.
Yes. That is the test of a complete ontology.
Can an agent safely act on it?No. It carries no ownership and no limits.
No. It says what is linked, not what is allowed.
Yes. Ownership and attachment say who may touch what.
Semantics is one layer. There are ten.
- 01Stable identity
The same piece keeps one id across rename, move and edit.
- 02Form
The shape of one piece.
- 03Hierarchy
Which piece contains which.
- 04Semantics
What it means: purpose, signature, summary, contracts.
- 05Pragmatics
What it is for in use.
- 06Relations
The typed edges: contains, calls, depends on, implements, mutates, observes.
- 07Dynamics
Behavior inside one boundary.
- 08Kinetics
Motion across boundaries.
- 09Claims
The laws the other layers are checked against.
- 10Attachment
Where an outside agent takes hold.