The Genie Suite Is Now a Platform, Ontology Included

Pull-quote: “Agents were failing at enterprise data the way tourists fail in a city without a map: intelligent, motivated, and lost. An ontology layer is the map, drawn before anyone asks for directions.”

Why this matters

The weakness of every enterprise agent to date has been the random walk: a question arrives, the agent opens a document, follows a link, calls a tool, reads another document, and loops until it either finds context or runs out of budget. Slow, expensive, inconsistent. At DAIS 2026, Databricks named this the context problem and shipped an answer with architecture behind it: Genie Ontology, a continuously updated knowledge graph of organizational context built before the question arrives, ranked by an algorithm called OntoRank across tables, documents, code, and the organizational graph itself. Databricks reported internal evaluations where ontology-grounded answers reached 84.5% accuracy on real data questions, against roughly 50% for leading coding agents — their numbers, announced at DAIS 2026, but directionally consistent with what anyone running agents against raw estates has observed.

The suite, sorted

Unity Catalog grew three semantic capabilities to feed the ontology — Business Glossary (preview coming soon), Domains (public preview), and Metrics (core features in public preview). That detail is the operating-model tell: the ontology is only as good as the governed semantics underneath it.

What an ontology layer fixes, and what it moves

   Without ontology                     With ontology
   ────────────────                     ─────────────
   Question ──► agent random-walks      Graph built continuously,
   the estate per query:                permissions applied, ranked:
   open doc ► follow link ►             Question ──► context retrieved
   call tool ► read ► loop...           in milliseconds ──► answer
   Slow, costly, inconsistent           Fast — and only as right
                                        as the semantics feeding it

The fix is real: context retrieval moves from minutes of live traversal to milliseconds of lookup. What it moves is the burden. The failure point is no longer the agent’s search; it is the glossary entry nobody wrote, the metric defined three ways, the domain boundary drawn wrong. We made this argument about Genie spaces before DAIS: treat them as data products with owners, curated scope, and certified answers. The ontology raises the stakes on exactly that discipline, because a curated semantic layer now feeds every agent in the suite, not one chat window.

What we would do with a client estate

Fund the semantics before the rollout: glossary, domains, and governed metric definitions are now agent infrastructure, and they need owners the way pipelines do. Pilot Genie One in one domain with a real question set and measure against it — the certification loop from the Genie-spaces playbook applies unchanged. Treat ZeroOps and the Ontology as previews: valuable to trial, wrong to depend on. And plan the last mile, because a governed answer that ends its life in a screenshot has lost its provenance. This is the seam our AlchemyLake platform occupies: it binds Genie spaces as governed sources for rendered deliverables, and its no-egress text tier runs on Foundation Model endpoints inside the workspace, so the chain from Unity Catalog to the finished artifact stays intact.

Closing

DAIS 2026 turned Genie from a natural-language BI feature into a platform: one context layer under a business interface, reusable agents, a coding agent, and an operations agent. The engineering is Databricks’ problem now. The semantics are yours. Estates that treat the glossary, the domains, and the metrics as governed products will get the 84.5%; estates that do not will get a faster random walk.