Databricks makes Genie Ontology generally available, claims 91% pass rate in Turing benchmark

Databricks has made Genie Ontology generally available. It is an agentic ontology that agents build from teams' everyday work and that people govern, so every agent shares one understanding of the business. According to Databricks, in a benchmark built with Turing, Genie with Genie Ontology passed 91% of enterprise task runs, 1.6 times the rate of the strongest frontier agent, while using less than half the tokens.

What shipped with GA

Databricks introduced Genie Ontology at its Data + AI Summit. The GA release adds:

• Genie Search, now generally available, for enterprise knowledge.
• Learned context extended to BI tools and organizational directories.
• Ontology Manager, for curating and governing what agents learn.
• Unity Catalog Domains and Pages, now generally available.

Databricks says thousands of customers already run Genie Agents in production with hand-curated context, including 7-Eleven, Virgin Atlantic and General Motors.

How it works

Genie Ontology combines two kinds of knowledge.

Learned knowledge. Agents extract business context from Databricks dashboards, SQL queries, notebooks and semantic models in tools such as Power BI. The context is stored as Ontology Snippets: reusable calculations, business rules and data relationships, updated continuously as the underlying sources change.

Modeled knowledge. Teams define and govern exact business meaning through Unity Catalog semantics. This includes Metric Views holding certified KPIs, Domains that organize knowledge by business area, and Pages that record authoritative definitions of concepts. Databricks cites recognized revenue and ARR as examples of definitions that should be modeled explicitly. Genie Ontology treats these as the source of truth.

When an agent asks a question, OntoRank, the authority-aware ranking algorithm, evaluates both kinds of knowledge. It weighs source authority, certification, the expertise and relevance of the people behind the knowledge, popularity and recency to resolve conflicting definitions. Existing user permissions are respected. Through Genie MCP, the same context is available to agents outside Databricks, such as ChatGPT, Claude and Microsoft Copilot.

Structured data enters Unity Catalog via Lakeflow Connect or is federated in place, so Genie can compute answers from live operational data.

Benchmark results

Databricks and Turing built a benchmark simulating a live Databricks workspace for a fictional UK energy supplier. It compared Genie, powered by Genie Ontology, with two unnamed frontier agents. Each agent answered every task five times, graded against a rubric defined before the runs began.

Agent	Pass rate	Avg wall time (per run)	Avg tokens (per run)
Genie	91%	100 seconds	432K
Agent A	45%	435 seconds	889K
Agent B	58%	307 seconds	1.05M

Databricks says Genie finished three to four times faster. It attributes the gap to context rather than SQL ability, since all three agents write SQL well. Genie used knowledge the Ontology had already learned, while the frontier agents had to rediscover it on every run and often returned answers that broke the relevant business rules. Databricks notes that the results reflect an established deployment where teams have accumulated that knowledge. The company designed and co-built the benchmark, and a separate research blog covers methodology and task walkthroughs.

What this means

The benchmark compares agents with and without accumulated business context, not raw model capability. Databricks reports that all three agents wrote SQL well, and that Genie's advantage came from already holding the relevant business rules. For teams evaluating agents on enterprise data, the relevant question is how business definitions are captured, ranked and governed, not only which model sits underneath.

The design also splits responsibility. Agents learn the long tail of team-specific logic from dashboards, queries and notebooks, while people model only the small set of definitions that must be exact. Genie MCP extends that shared context to agents running outside Databricks. Because the results come from an established deployment with accumulated knowledge, Databricks suggests customers run the same evaluation in their own workspaces through their account teams.