Google Makes Knowledge Catalog Semantic Support Generally Available

Google said AI agents (software that answers business questions) grounded in Knowledge Catalog were nearly four times more accurate in internal benchmarks, with up to 33% lower latency and up to 75% fewer tokens. Latency means response time; tokens are units of text processed by the model. Google compared agents using catalogued business context with agents answering without it.

Google made semantic model support generally available on November 14, 2026. The update adds semantic models (machine-readable definitions of business metrics and relationships) from six analytics and data platforms to a shared context graph (a map of data objects and their links).

That matters because enterprise AI can produce fluent language and still return the wrong number when systems rely on conflicting or unverified definitions, Google said. Google said Knowledge Catalog gives agents “a single source of business context drawn from tools already used by analysts and data teams.”

What changed on November 14?

Google’s release says Knowledge Catalog now ingests semantic models and metric definitions from Looker LookML, BigQuery Graph, dbt MetricFlow, Snowflake Semantic Views, Databricks Metric Views and SAP Core Schema Notation.

Google says the catalog links those definitions with schema, lineage, data profiles and business glossary metadata. The process does not move data or require companies to rewrite business logic. Instead, agents can use definitions already maintained inside reporting, analytics and data engineering tools.

Google designed the update for conventional business reporting as well as AI-assisted analysis. It presents the catalog as shared context for agents while definitions remain embedded in the tools used by analysts and data teams. The expanded semantic-model ingestion is generally available; Knowledge Studio remains in preview.

The status difference has a practical effect: teams can use the expanded catalog ingestion, while the interface for inspecting and packaging context remains a preview product.

How does the context graph work?

Google maps four types of object as nodes: tables, semantic models, glossary terms and data products. It connects them through lineage (a record of how data moves and changes), associations, schema keys and joins.

Google says the graph also incorporates join paths declared in semantic models and joins inferred from query history. That structure gives an AI system a route from a business concept to a formula and then to the underlying data sources.

The architecture reduces the risk of agents drawing on staging tables (temporary holding areas), stale data or inconsistent definitions, according to Google. The graph can connect semantic definitions with data quality scores, freshness signals and lineage information. Systems can then assess whether a source merits trust before generating a query.

Google describes a specific failure mode: an agent can write fluent prose while mapping a business concept to the wrong formula. The graph adds filters, lineage, synonyms and quality signals before query generation. It does not make conflicting definitions disappear.

Which analytics platforms are supported?

Google documents a different extraction pattern for each supported platform:

• Looker LookML: Live synchronization from Looker Core indexes models, explores, views and dashboards. Google said the service captures compiled metric SQL (database queries), join conditions, mandatory filters and lineage from source tables through dashboards.
• BigQuery Graph: The catalog indexes property-graph definitions (structures built from connected nodes and edges), node and edge tables, graph measures and business synonyms. The aim is to let agents reason across connected entities through graph traversals rather than multiple table self-joins.
• dbt MetricFlow: The catalog links source objects with dbt models and metrics, alongside data quality test results, filters, saved queries, freshness service-level agreements and downstream exposures.
• Snowflake Semantic Views: Google ingests dimensions, facts, metric formulas, relationships and synonyms, plus Data Metric Function quality scorecards and masking tags (labels for controlling data visibility).
• Databricks Metric Views: The service ingests metric view definitions, formulas, filters and join cardinalities (how records match across tables) from Unity Catalog, together with table certification badges and monitoring results.
• SAP Core Schema Notation: In SAP environments, Google captures business labels, modelling patterns, default aggregations, currency and unit bindings, join associations and data sensitivity flags when SAP Business Data Cloud products federate through Google Cloud Lakehouse.

How does Google handle conflicting metrics?

Google says Knowledge Catalog keeps both definitions visible when tools define the same metric differently, and shows their provenance so data stewards can reconcile them. The software does not silently overwrite competing definitions. Google cited revenue, margin and customer counts as figures that can vary with filters, joins or exclusions.

Google’s release states: “Overlapping definitions remain side by side, each labelled with its source, giving data stewards the full picture they need to settle on a single version.” The decision on which version becomes authoritative therefore remains with data stewards, not an automatic overwrite.

That policy matters because a shared metric name can conceal different filters, joins or exclusions behind apparently matching figures. The catalog makes the disagreement visible, but Google still assigns the final selection to the people responsible for business definitions.

Where does Gemini fit?

Google frames the release as a governance effort to tie AI outputs more closely to corporate data controls. The company wants agents to reuse definitions embedded in reporting, analytics and data engineering tools, rather than maintain a separate layer of business logic for AI.

Knowledge Catalog can combine structured semantic models with context extracted from unstructured materials, including operational playbooks, meeting notes, chat threads and code. Google said Gemini drives this “self-learning enrichment,” which the company intends to keep context current as definitions change.

What does Knowledge Studio add?

Knowledge Studio remains in preview. Google describes it as a control point for exploring and managing the context graph. Users can inspect semantic definitions with natural-language prompts and package validated assets into reusable context sets for specific use cases or business domains.

The preview covers the management interface, while the expanded semantic-model ingestion is already generally available. That leaves Knowledge Catalog operating at two stages: automated context collection in production-facing systems, with human-directed inspection and packaging still under preview.

Why does this matter for enterprise AI?

Google positions Knowledge Catalog as a control layer between corporate data estates (an organization’s collection of data systems) and the AI tools that query them. Its release identifies the central enterprise weakness of generative AI as weak alignment with internal business definitions, rather than language generation itself.

For data teams, the stakes are clear: conflicting definitions must remain inspectable while agents run against them, and stewards must resolve the differences without rebuilding analytics infrastructure. Google’s approach reuses existing tools instead of asking companies to create a separate metric layer for AI.

The catalog maps labels, formulas, filters and joins, while data stewards still determine which definition governs. Google’s division preserves analyst workflows without treating an agent as the final authority on business meaning. The practical test for enterprise AI deployments is whether teams can turn those side-by-side versions into governed definitions while Gemini keeps enrichment current. Google’s release leaves reconciliation with the stewards responsible for business metrics.