What happened
On October 29, 2026, Google Cloud announced general availability of expanded semantic-model support in Knowledge Catalog. Knowledge Catalog is presented as a context graph that joins semantic models and metric definitions with schema, lineage, data profiles, business glossaries, and other metadata. The announcement focuses on a familiar failure mode for analytical agents: an assistant can return a polished chart yet still disagree with finance because it used an unvalidated table, included cancelled orders, or applied a different definition of a metric.
The GA capability natively ingests semantic models and definitions from Looker LookML, BigQuery Graph, dbt MetricFlow, Snowflake Semantic Views through Horizon Catalog, Databricks Metric Views through Unity Catalog, and SAP Core Schema Notation through SAP Business Data Cloud. Google says this does not require moving data or rewriting business logic in a new language. The resulting graph links semantic material with the surrounding governance and usage information, allowing agents to move from a business concept to a formula and declared or observed join paths.
The source describes different details retained for each source. LookML ingestion can include models, explores, views, dashboards, compiled metric SQL, join conditions, mandatory filters, and lineage. dbt MetricFlow can bring metrics alongside data-quality test results, filters, saved queries, freshness service-level agreements, and downstream exposures. Snowflake and Databricks inputs can add semantic definitions as well as quality, masking, certification, or profiling signals. SAP input can include labels, modeling patterns, aggregations, currency and unit bindings, join associations, and sensitivity flags.
A significant design choice is that overlapping definitions are not silently merged. If two systems define the same metric differently, both remain in the graph with their provenance. That gives data stewards an opportunity to reconcile a conflict before an agent presents an answer as authoritative. Google says the graph can ground Gemini agents, BigQuery Conversational Analytics, and custom agents in the metrics, filters, and joins already used by analysts. It also describes a Gemini-powered self-learning enrichment feature that extracts relevant context from unstructured sources such as playbooks, meeting notes, chat threads, and code, and turns it into governed context sets as definitions change.
Google reports internal benchmark results of nearly four times higher agent accuracy, up to 33% lower latency, and up to 75% fewer tokens when agents are grounded in Knowledge Catalog. Those figures are Google’s own benchmarks, not an independently reported evaluation, but they indicate the efficiency and reliability claims attached to the release.
What the archive adds
An earlier Telegram post described Knowledge Catalog as a shared enterprise-context layer for data, semantics, and business logic. This is a new fact on that known product story: the multi-system semantic-model ingestion is now generally available, with named inputs from Looker, dbt, Snowflake, Databricks, SAP, and BigQuery.
Trend signal
This supports Company memory needs write and correction loops. The release is not simply a larger searchable document store: it combines governed semantic definitions, provenance, lineage, freshness and quality signals, while Google’s self-learning enrichment is intended to update reusable context as the underlying business changes. The important implementation question is whether those correction and stewardship loops work reliably in a real organization.
Why it matters
Data agents fail expensively when they answer a well-formed question against the wrong definition of the business. A generic model can translate “net revenue for top enterprise accounts” into plausible SQL, but it cannot infer which cancellations count, which tables are certified, or which join has been approved. Those decisions often already exist inside separate analytics products.
Knowledge Catalog’s approach is to make those existing decisions available as context rather than requiring every agent team to reproduce them in prompts, retrieval pipelines, or hand-written tools. Keeping conflicting definitions visible is also important. A system that picks one silently can make an answer look more certain than the organization’s data governance actually allows. Teams using several of the supported platforms can treat semantic governance as an agent dependency, not an afterthought to query generation.
Limits and uncertainty
The announcement is a product description from Google Cloud. Its reported improvements in accuracy, latency, and token use are based on internal benchmarks; the supplied material does not describe the benchmark tasks, baselines, datasets, model configurations, or error rates. It also does not establish how much implementation work is needed to connect each source, resolve conflicting metrics, set access controls, or operate stewardship workflows.
A unified context graph can expose disagreements, but it does not itself decide which business definition should prevail. The source says that data stewards can reconcile conflicts, not that reconciliation is automatic. It also does not show independent customer results or compare the service with alternative metadata and semantic-governance architectures.
Skill to build now
Inventory one high-value agent question before adding another model or SQL prompt. For each metric in the answer, identify its approved definition, source system, mandatory filters, permitted join path, data-quality check, freshness expectation, and owner. Then test whether your agent can retrieve those items separately and whether it signals a conflict rather than choosing quietly. This produces a concrete semantic contract that can later map to a catalog, graph, or internal tool layer.
Reader takeaway
Google Cloud’s GA update turns Knowledge Catalog into a broader ingestion point for semantic definitions already held in major analytics tools. The durable lesson for data-agent teams is to ground answers in governed metrics, lineage, and quality signals. Better query generation cannot compensate for an agent that does not know which version of a business definition to trust.