Databricks Sets a Six-Layer Path for Genie Ontology, Starting With One Domain

Databricks has published a six-layer operating path for Genie Ontology that begins with clean, reconciled data and ends with continuous evaluation. The company’s practical advice is deliberately narrow: start with one business domain, then expand as definitions, access controls and trusted context mature.

The guide frames an ontology as more than a semantic model. Genie Ontology combines explicitly modeled business semantics with context inferred from governed tables, queries, dashboards, notebooks and other supported assets. Databricks says that context is ranked by authority and relevance and filtered through permissions before it reaches Genie at answer time.

The work starts before the agent asks a question

Layer 0 is a data-modeling correction. Databricks recommends durable business-process models with a clear grain, reusable dimensions and a single reconciled identity for each real-world entity. The goal is to avoid giving an agent tables with conflicting concepts, mixed levels of detail or several IDs for the same customer.

The next layer is descriptive rather than logical: table descriptions, column comments and structured tags for attributes such as sensitivity, ownership and business function. For large estates, Databricks points to its dbxmetagen accelerator, which can draft descriptions, flag sensitive data and propose classifications. A human must approve changes before they are written to Unity Catalog.

Three controls Databricks puts at the center

1. Declare primary and foreign keys so agents have defined table relationships rather than inferred join paths. Those declarations are informational, not enforced, so they still require upkeep.
2. Define critical KPIs in Metric Views, where governed measures and dimensions are expressed as code and aggregations are resolved when a query runs.
3. Use Domains and Pages to scope business context and attach authoritative assets to shared terms and definitions.

Pages give human definitions priority

The guide’s curation layer relies on Unity Catalog Pages, a Beta feature for governed definitions of business concepts. When Genie One finds a matching Page, it prioritizes that human-modeled definition over automatically inferred context and cites the Page in its answer. Published Pages are available across Genie One conversations, while drafts are limited to the Page owner’s conversations.

Governance creates a sharper boundary

The operating model also leaves important limits for adopters to manage. Because key declarations are not enforced, a declared relationship can still become stale. And Pages inherit access from their domain or subdomain, rather than receiving permissions directly. Databricks also warns that Page data is stored as plain text that may be replicated globally and does not support customer-managed-key encryption; it says Page names, descriptions and bodies should not contain PII or regulated data.

The final layer is evaluation. As metrics change, tables age and products are introduced, Databricks recommends measuring answer quality to identify weak context and improve it. The unresolved operational question is how each organization will set those quality thresholds and assign owners for the business definitions that agents rely on.