What an ontology is in Cognee
An ontology is the structure your memory is sorted into. It has three parts:
Every entity Cognee extracts (a person, a ticket, an invoice) gets the most specific entity type that fits it.
Default ontology vs. auto-generated ontology
Default (free). Cognee extracts entities with generic types such as Person, Organization, Location, or Concept. That works for any dataset, but it does not know that, in your company, a customer and an employee are different kinds of person, or that incidents and incident records are the same thing. Auto-generated (Premium). Cognee reads your dataset and works out the entity types your company actually talks about, with definitions, a hierarchy, and the relations between them. Your entities are then sorted into these sharper types. On the Graph page the result appears as extra domains and sub-domains, and the entity types and entities that are new or changed are highlighted.Generate an ontology
The Premium ontology view on the Graph page, with its generate action and the ontology chat editor, is rolling out with the November 2026 launch, so exact labels may differ slightly from this page. Until it reaches your workspace, the editor is available through the API.
1
Add data first
Auto-ontology works from what is already in a dataset, so add data and let it finish processing. The richer the dataset, the better the taxonomy.
2
Open the Graph page
Open Graph in the sidebar and choose the dataset.
3
Start the generation
Switch to the Premium ontology view and start the generation. You need an active Premium Ontology plan; see Who can use it.
4
Keep working while it runs
Generation runs in the background. It takes several times longer than a normal ingestion (roughly four times as long), so you do not need to wait on the page. When it finishes, the Premium ontology view shows the new domains, entity types, and the entities that moved into them.
Refine it with the ontology chat editor
The ontology chat editor lets you change the ontology by describing what you want, instead of editing a schema file. Open it from the Premium ontology view, or from the card of an entity type to make a change about that type.1
Describe the change
Write the change in your own words. A request can hold several changes, up to 4,000 characters.
2
Review the proposal
Cognee reads your request and turns it into concrete changes, each with its impact (for example, how many entities it affects). Nothing changes yet. A change Cognee could not understand, or one that cannot be applied, is marked as such.
3
Apply
Confirm to apply the proposal. All changes in it are applied together, or none of them are.
- Merge incidents into incident records.
- Rename Client to Customer.
- Split Event into Customer meeting and Internal meeting.
- Move Invoice into Finance.
- Add a new type Approver: the person who signs off a credit request.
- Customers own contracts. (adds a relation)
- Employees do not report to customers. (removes a relation)
Renames, merges, moves, new types and relation changes are stored as decisions on top of the graph, not written into it, so they survive rebuilds: when the graph is built again, the same decisions apply. A split is the exception: it re-types the entities in the graph itself. A dataset can hold up to 200 added entity types and 1,000 relation changes.
Reading a request and splitting a type call the language model, so they use workspace credits like generation does.
How it improves retrieval
Agents and search read the graph through its entity types. With an ontology built from your data:- Questions land on the right category (customers vs. employees, incidents vs. incident records), so fewer irrelevant entities reach the agent’s context.
- Entities that were scattered across generic types are grouped, so “how do credit approvals usually work” can be answered from one entity type and its relations instead of dozens of loose chunks.
- Relations carry meaning (performs, requires, blocks) instead of a vague “related to”, so multi-step questions follow the right path.
Who can use it
A Premium Ontology subscription belongs to a workspace owner and makes every workspace that person owns Premium; members of those workspaces get it at no cost. Only an owner can subscribe; members see “Only a workspace owner can subscribe.” Subscribe from Premium in the top bar or on the Billing page; see Premium Ontology billing.
When Premium ends
If you cancel, Premium stays active until the end of the paid period. After that, your workspaces lose:- Automatically generated ontologies: new data is no longer modelled into ontologies for you.
- The ontology chat editor: you can no longer refine ontologies by chatting with the editor.
- Extra domains and sharper types: your knowledge graph goes back to the standard ontology.
API
The ontology chat editor is also available on the tenant API, under/api/v1/datasets/{dataset_id}/knowledge-map/edits. Making a change needs an active Premium Ontology plan and write permission on the dataset; reading the changes in force needs only read permission and works on any plan.