Method · Ontology
AI doesn't need more documents. It needs a model of your business.
An ontology organises what actually exists in your company (customers, orders, parts, machines), how those things relate and what may happen to them into a model AI can query precisely and act on safely. It's the missing layer between LLMs and your production systems.
What an ontology is
Three definitions describe an entire enterprise.
An ontology isn't a document or a picture. It's a typed, rule-bound, executable business model. Data keeps flowing in from your systems, and people and agents use it through the same interface.
Order SO-8812
- statusconfirmed
- amount¥186,400
- dueNov 6
One question, four answers
When the question needs facts, the difference shows.
Ask four architectures the same real supply-chain question. All of them will “answer”. Only one gives you something you can act on.
RAG
similarity searchChunks documents into vectors and hands the model the passages most similar to the question.
GraphRAG
inferred relationsUses a model to extract entities and relations from documents into a graph, then answers with community summaries.
LLM Wiki
updated 3 weeks agoA model compiles sources into markdown pages and maintains them; agents read the pages.
Ontoxiom ontology
live · authorisedTraverses typed relations synced live from ERP, with every step authorised by the Governor.
Side by side
They solve different problems.
| RAG | GraphRAG | LLM Wiki | Ontoxiom ontology | |
|---|---|---|---|---|
| Source of knowledge | Document chunks | Graph extracted from documents | Model-written pages | Systems of record + documents, one model |
| Freshness | Depends on re-indexing | Depends on graph rebuilds | Depends on the model rewriting | Continuously synced per integration |
| Exact counts & lists | No | Limited | No | Yes: query, traverse, aggregate |
| Where relations come from | None | Inferred by a model | Links between pages | Declared, with direction and cardinality |
| Permission granularity | Document | Document | Page | Object, property, relation, action |
| Can act on the business | No | No | No | Yes: check → approve → execute |
| Auditability | Retrieval logs | Retrieval logs | Page revision history | Every read, write and decision recorded |
| Best for | Policy & manual Q&A | Exploring themes in a corpus | Personal and team knowledge | Production work that needs facts and actions |
FullyPartlyNo
Why it matters
Production AI needs four things documents can't give.
Precise
“Which customers dropped more than 20% this month?” needs a query and an aggregate, not similarity search. An ontology returns answers you can enumerate and check.
Live
Order status changes by the minute. Data Pipeline keeps the ontology synced with your systems, so the model reads today's facts rather than last month's summary.
Actionable
Going from answering to moving a process forward needs defined actions, preconditions and approvals. Only an ontology keeps actions and facts in one model.
Governable
Who may see the salary field, which action needs approval: permissions must live on objects and properties. Document-level access is too coarse for an enterprise.
Not a replacement for RAG. A factual foundation underneath it.
Document retrieval is still useful for policies, manuals and know-how. In Ontoxiom, agents can use documents and the ontology together: explanations come from documents, while facts and actions come from the ontology, and both answer to the same permissions.
FAQ
Questions about ontologies
An LLM Wiki has a model compile sources into markdown pages it maintains itself; agents read those pages to answer questions. It's great for capturing personal or team know-how.
But a wiki is a model-written summary. It can be stale, it can't count or enumerate precisely, it has no field-level permissions and it can't execute business actions. Agents in Harness work on typed objects continuously synced from your systems, so answers are exact down to the individual order. Every read and write passes the Governor, and actions go through checks, approvals and a full audit trail.
They can coexist: the wiki holds the “how”, the ontology holds the “what” and the “do”.
Turn your business into a world AI can understand.
Bring one concrete business problem. We'll tell you honestly whether it makes a good first use case.