Genie Ontology and Unity Catalog Semantics: The context layer arrives on the platform

As of: August 2026. Feature and preview details per Databricks announcements at the Data + AI Summit 2026 and the public docs.

Two teams ask the same lakehouse the same question: "What was the active customer count last quarter?" Both get a fluent answer. Both answers sound authoritative. They differ by 18 percent.

No model error. No data error in the narrow sense. Finance counts accounts with a paid invoice. Product counts sessions over 30 seconds on a core feature. Both definitions existed somewhere: in a Genie Space, in a dashboard comment, in a Slack reply from eight months ago. The platform had data. The organization had no shared context.

This is exactly the diagnosis Databricks put center stage at the Data + AI Summit 2026. CEO Ali Ghodsi put it into words we've used as a guiding theme since our post Making data AI-ready: AI doesn't have an intelligence problem, it has a context problem. The platform's answer isn't "a better system prompt." It's Unity Catalog Semantics and Genie Ontology – curated semantics as the foundation, with a living knowledge graph on top that feeds agents, BI, and apps with the same meaning.

This article puts the announcements into solution-architecture terms. It shows which building blocks do what, how they fit our context layer, and where the line runs between better retrieval and a decision you can still defend tomorrow.

1. Why longer prompts make the problem bigger

Most enterprise teams solve context today the way people solved SQL before the semantic layer: by copying it.

In every Genie Space, every agent, every RAG bot, you find variants of the same rules:

• What "revenue" means,
• which table is authoritative,
• which synonyms apply,
• which exceptions the business unit has "always" known about.

Databricks reports more than a million Genie Spaces. That's a success for the self-service idea – and, at the same time, a structural risk. Every space can carry its own truth. If the author leaves the company or the business changes the definition, the prompt quietly ages. The model keeps answering confidently. No one gets an error. You get drift.

That's the same logic we criticize under Evidence-first AI: linguistic plausibility is not yet a solid basis for decisions. An assistant that mixes meaning, calculation, access, and phrasing into a single step produces answers that sound backed by evidence – without anyone being able to say which definition and which scope were actually binding.

The historical parallel helps here: the same metric used to live as slightly different SQL in fifty dashboards. The solution wasn't discipline around copy-paste. It was a semantic layer: define once, consume everywhere. Genie Ontology and Unity Catalog Semantics are an attempt to make the same leap for AI context – with the difference that Databricks wants to automatically learn part of that context from the organization, rather than maintaining all of it by hand.

2. What Databricks announced – and how the pieces work together

Three product lines belong together, even if marketing sells them separately:

Official starting points: Introducing Genie One, Genie Ontology, and Genie Agents, Unified context: The missing layer for enterprise AI coworkers, Unity Catalog Semantics, and What's new with Unity Catalog at DAIS 2026.

The architectural logic is clear:

1. User-defined semantics in Unity Catalog provide the high-authority inputs (certified terms and metrics, business-scoped domains).
2. Genie Ontology additionally learns from usage and other sources and connects it all into a business-aware map.
3. Agents and BI consume the same context – including permissions from Unity Catalog.

Databricks calls this "decision-ready context": not merely data in one place, but a shared map where numbers, definitions, and relationships fit together. Without this map, forecast calls and steering committees get bogged down in reconstructing which pipeline, which usage signal, and which account history was actually meant.

3. Unity Catalog Semantics: Glossary, Metrics, Domains

3.1 Business Glossary – Meaning outside the model

The Glossary in Unity Catalog is meant to hold binding concepts, terms, and taxonomies – importable from existing definitions, linkable to data assets and to each other. Genie Code can propose drafts and flag drift; ownership and approval stay with the team. Preview was announced as "coming soon" at DAIS 2026.

From a business perspective, this is layer one of our context layer: meaning must be defined outside the model. An LLM may offer suggestions. It may not decide whether "active customer" means the Finance view or the Product view. This is exactly where many AI demos fail: they silently pick one variant and sell it as clarity.

3.2 Metrics / Metric Views – Calculation as a platform object

Metric Views are the hard edge of semantics. A metric is modeled once as a reusable object: sources, dimension fields, measures, optional joins, synonyms, and formatting for agents. Unlike classic views, aggregation and grouping aren't frozen once and for all. The definition stays declarative; the engine generates the appropriate calculation at runtime.

At DAIS 2026, Databricks extended the modeling: multi-fact relationships, level-of-detail calculations, parameterized metrics, better window measures, materialization for faster dashboard and agent queries, and import paths from Power BI and Tableau (beta). Metrics are also positioned as open semantics – with an Apache Spark / Unity Catalog OSS path and OSI readiness.

This is layer two: calculation needs a binding logic. The model may explain the value. It shouldn't reconstruct it from schema and gut feeling. The same question in the SQL editor, an AI/BI dashboard, and a Genie agent must hit the same metric – otherwise you've just copied prompt drift into three surfaces instead of one.

3.3 Domains – Scope before context

Domains (public preview) cut data and AI assets into business-aligned categories. Agents and people don't get the whole catalog, but a business-bounded horizon, with certification and stewardship signals.

That sounds organizational. It's architectural. An agent allowed to see "everything" that technically exists in the metastore becomes uncertain or overreaching. Domains are the entry point into layer four – agent context – even before permissions kick in at a detailed level: which slice of the world is even relevant to this question?

4. Genie Ontology: Living graph, OntoRank, benchmarks

Genie Ontology is the layer Databricks markets as the unified context layer. It extracts knowledge snippets from tables, queries, dashboards, pipelines, and connected apps, and organizes terms, metrics, entities, and relationships into a continuously updated graph. Genie One and Genie Agents are meant to know, based on this, where to look, what to trust, and how the organization actually uses its data.

4.1 OntoRank: Authority instead of similarity

The decisive mechanism, per reporting, is called OntoRank – a ranking analogous to PageRank, applied to business definitions across heterogeneous asset types. Databricks and keynote coverage cite signals including:

• provenance and relative authority of the source or author,
• breadth of usage (how often teams and queries access it),
• linkage to certified and widely used assets,
• recency.

Genie then preferentially answers from the sources with the highest weight – and only with content the requesting person is allowed to see. Permissions aren't "requested in the prompt," but enforced via source-native ACLs or Unity Catalog.

This is the qualitative leap over classic vector RAG: RAG asks What resembles the question? OntoRank asks What counts as authoritative in this organization? Similarity and authority are different currencies. Whoever only has similarity finds the right wiki paragraph from 2021. Whoever has authority prefers the certified Metric View from Controlling.

4.2 What Databricks claims about performance

In the announcement blog, Databricks cites an internal benchmark (28 real-world data-analysis questions, June 2026): Genie with Ontology answered 84.5% of questions correctly on the first attempt; the strongest general-purpose coding agent reached 52.4%, the weakest 25%, with lower latency at the same time (roughly 2× faster than the strongest coding agent). These are vendor figures, not an independent audit – but the direction of the thesis is clear: context beats raw intelligence once the question becomes company-specific.

4.3 What Genie Ontology is not

This is where marketing and architecture part ways. A Databricks employee clarified in the community in July/August 2026: Genie Ontology is not, today, a full formal, queryable ontology in the sense of OWL/SPARQL. It automatically ranks and resolves conflicts between snippets. Citations show which sources fed a specific answer. A dedicated export, snapshot, or changelog of all newly "learned" edges is not yet publicly documented.

That's not a minor detail for regulated organizations. A semantic layer that evolves without a reviewable diff remains an adoption blocker for many governance boards – even if every individual answer comes with citations. Analysts at InfoWorld raise the same reservation: ontologies improve context, they don't guarantee correct execution; without ownership, lineage, and cleaned-up definitions, they just accelerate the existing chaos.

Our reading: Genie Ontology is a capable context layer – not a substitute for business sign-off. Automatic induction produces suggestions and connections. Bindingness only emerges once Glossary and Metrics have ownership and conflicts are handled visibly.

5. Genie One, Genie Agents, and the path into the tools

Genie started as conversational analytics in AI/BI. Genie One is the next step: a data-smart AI coworker that, via Lakehouse Federation, Lakeflow Connect, and integrations (Slack, Teams, Gmail, Mobile, MCP), doesn't just answer but continues the work – briefings, document updates, schedules, skills.

Genie Agents extend the Genie Spaces model: curated, domain-specific chat experiences become shareable agents that use structured and unstructured sources and can act via MCP and workflows. Creation from a prompt, scoping, benchmarking, sharing.

For the architecture, the surface matters less than the invariant: many agents, one context. Without shared semantics and ontology, you reproduce the 18-percent conflict – just this time between a revenue agent and a churn agent, both in production.

Governance deliberately sits in Unity Catalog and the Unity AI Gateway: models, agents, tools, and MCP services under a single runtime governance; contextual service policies (beta) for allow/deny/approval of actions; budgets and tracing. This is the platform-side answer to "governance in the prompt."

6. Mapping to beyond chaotic analytics' context layer

We describe four layers between raw data and a trustworthy AI answer – later extended by the data platform as the foundation. On Databricks, they map as follows:

The platform increasingly delivers three of these layers natively. The third remains demanding: Genie Ontology learns relationships and context; for questions that run across multiple hops and are subject to an evidentiary requirement, a learned context graph often isn't enough. Then you need modeled edges with ownership – the subject of Knowledge graphs in the enterprise, including the Neo4j connector and the GraphRAG reference on the Mosaic AI Agent Framework.

And the evidence-first rule remains unchanged: meaning is not calculation. Calculation is not access. Access is not phrasing. The LLM comes at the end of the chain. Details in the post The LLM may answer. It may not decide what's true..

7. Decision landscape: RAG, Metrics, Ontology, Graph

Not every question deserves the same stack. The choice of tool follows the type of question – not the feature catalog:

Genie Ontology replaces neither Metric Views nor a formally modeled graph. It reduces the cost of sharing context across spaces and agents – and it prioritizes what the organization already signals as trustworthy. Where popularity and business correctness diverge, you still need an owner and sign-off, not just a ranking.

8. Critical assessment: What improves – and what remains

What improves

• Fewer prompt silos between spaces and agents.
• Certified metrics and glossary entries as high-authority anchors.
• Permission-aware answers instead of "please only use allowed data."
• MCP-addressable semantics: context as infrastructure, not a one-off instruction.
• Citations per answer – a step toward traceability.

What isn't automatically solved

• Incorrect but popular definitions (popular ≠ correct).
• Conflicts between two highly authoritative sources – the algorithm isn't publicly specified.
• Missing ownership and unresolved domain boundaries.
• Formal action guardrails before risky write actions (AI Gateway policies help here – plus business rules outside the model).
• Auditability of ontology evolution (changelog/diff still thinly documented).
• Relationship questions with an evidentiary requirement, absent an explicit graph model.

So the realistic thesis isn't: "Ontology cures hallucinations." It's: Unity Catalog Semantics and Genie Ontology move parts of agent logic out of probabilistic prompts and into platform-native, permission-bound context objects. That's significant. It's not a universal truth machine.

The same standard that a board decision brief demands of its factual basis applies here: Does the same question lead to the same number next week – and can someone in the room explain the derivation?

9. A realistic rollout model

Not "ontology for the entire catalog." Instead, a guided build-up that enforces ownership:

1. Three to five steering metrics with the highest risk of dispute or liability, modeled as Metric Views – including synonyms and agent metadata.
2. Glossary for exactly these terms: definition, scope, example, owner, links to assets.
3. Cut a Domain around the affected business unit; set certification signals.
4. Build Genie One / a first Genie Agent on top of this and benchmark answers against the certified assets.
5. Only then expand: more domains, docs, connected apps, MCP consumers.
6. For relationship questions with an evidentiary requirement, decide separately: does ontology context suffice, or is a modeled graph required?

This creates compound interest instead of space archaeology. Every new metric and every glossary entry improves all agents – not just the one bot someone wrote the instruction into.

Sources and further reading

• Databricks: Introducing Genie One, Genie Ontology, and Genie Agents
• Databricks: Unified context: The missing layer for enterprise AI coworkers
• Databricks: What's new with Unity Catalog at Data + AI Summit 2026
• Databricks: Unity Catalog Semantics · Metric Views Docs
• Context: InfoWorld – From RAG to ontology
• beyond chaotic analytics: Context layer whitepaper · Clarity Audit

Revision note: First version August 2026. Preview and GA status may change; check current Databricks docs and your account team before making production decisions.