Sulat.com / Writing on AIMonday, November 16, 2026
Agents/16 November 2026

Google Cloud’s Knowledge Catalog learns missing business context from agent use, in preview

Google Cloud says self-learning enrichment and Knowledge Studio are in preview in Knowledge Catalog. Its claim of nearly 4x more accurate agents comes from internal benchmarks.

A glossy three-dimensional Google Cloud logo wrapped in a glowing blue network of documents, tables and databases, sending a beam of light to a small glass robot.

At a glance

  • Google Cloud says self-learning enrichment and a new Knowledge Studio are in preview in Knowledge Catalog, its context engine that grounds AI agents in business definitions, data and documents.
  • Google says enrichment reads Cloud Storage, Google Workspace, GitHub, Confluence, SharePoint and agent interaction traces, then proposes definitions, joins and instructions with source citations and a confidence score.
  • Admins set the confidence score at which changes publish automatically, Google says, with 85% as its example; lower-scoring changes and deletions go to a data steward for approval.
  • Google says Knowledge Catalog makes agents nearly 4x more accurate, with up to 33% lower latency and up to 75% fewer tokens, based on internal benchmarks the three posts don't describe.
  • Knowledge Studio's docs say the preview doesn't cover Assured Workloads regimes such as FedRAMP and HIPAA, data residency commitments, customer-managed encryption keys or VPC Service Controls.

Google Cloud has put self-learning enrichment and a new Knowledge Studio into preview in Knowledge Catalog, the service it calls its “universal context engine” for AI agents. Google says grounding agents in the catalog makes them nearly 4x more accurate, with up to 33% lower latency and up to 75% fewer tokens, based on its internal benchmarks.

Google announced the preview on 22 October and repeated it on 4 November in an Agentic Data Cloud roundup by Sirish Chandrasekaran, VP of product strategy and growth for Data Cloud, and Raj Pai, VP of product management. Knowledge Catalog was called Dataplex Universal Catalog until 10 April 2026, the release notes say. The API, client library, CLI and IAM names didn’t change.

What Knowledge Catalog does

Knowledge Catalog collects metadata, business definitions and relationships across a company’s data into a context graph. In Google’s description, tables, semantic models, glossary terms and data products are the nodes. Lineage, schema keys, joins mined from real query history and the join paths declared in semantic models are the edges. A semantic model is a company’s agreed definition of a metric, such as net revenue. Agents read the graph through the Model Context Protocol (MCP, the open standard that lets agents call outside tools) or through the catalog’s APIs, the docs say.

Flow diagram: data and models plus docs and code feed Knowledge Catalog's context graph, which AI agents query through MCP or an API.
Knowledge Catalog sits between a company’s data and documents and the agents that query them.

Semantic model support is generally available. Google says the catalog ingests models and metric definitions from Looker LookML, BigQuery Graph, dbt MetricFlow, Snowflake Semantic Views, Databricks Metric Views and SAP Core Schema Notation, without moving the data.

What’s in preview

Self-learning enrichment is the main addition. Google’s 22 October post says a Gemini-powered enrichment task connects to Cloud Storage, Google Workspace, GitHub, Confluence and SharePoint, and the 4 November post adds Jira. It proposes business definitions, table and column descriptions, join relationships, golden queries (vetted example queries) and task-specific agent instructions. Each proposal carries source citations and a confidence score.

It also learns from use. Enrichment can read interaction traces from Gemini agents, Conversational Analytics in BigQuery or custom agents, with personal data redacted, and flag where agents struggled: ambiguous column lookups, ungrounded metrics, missing definitions or users rephrasing the same question.

Google says admins can switch auto-publishing on or off and set the confidence score at which a change is approved automatically, with 85% as its example. Anything below the threshold, and sensitive updates such as deletions, becomes a change request for a data steward to approve, edit or reject. Every change is logged and can be reverted, Google says, and Google never uses customer data to train its models.

Four steps: agents and users ask, gaps get spotted, a fix is proposed with a confidence score, a steward approves or it auto-publishes.
A proposed fix passes a confidence score and a steward before agents see it.

Knowledge Studio is the console for that review. Google says context engineers and data stewards use it to explore the graph in natural language, see a visual map of their data estate and approve proposed changes with versioning and rollback. The docs say enrichment tasks are managed in the Google Cloud console, and REST APIs and gcloud commands for managing them aren’t available in preview.

Context sets are curated subsets of the graph for a domain, team or agent. Google’s example binds a sales agent to a Sales set and a finance assistant to a Finance set, because “net revenue” can mean different things to each group. Access to each set is controlled with IAM, and the docs say a set doesn’t bypass permissions on the underlying tables. Sets build on the Open Knowledge Format (OKF), so a bundle can be imported and exported.

Google also says Knowledge Catalog is now on by default in Gemini, in preview. When a user connects BigQuery, AlloyDB, Cloud SQL or Spanner, it resolves business terms, finds trusted tables and works out join paths before Gemini writes the SQL, under that user’s permissions.

The numbers

Google’s three figures appear in its 22 October, 29 October and 4 November posts, each tied to “internal benchmarks”. The posts don’t describe the questions, the baseline agent or the models used.

Three claims from Google's internal benchmarks: nearly 4x more accurate, up to 33% lower latency, up to 75% fewer tokens.
All three figures come from Google’s own internal benchmarks.
PieceStageWhere Google says so
Semantic model supportGenerally available29 October blog post
Self-learning enrichmentPreview22 October blog post
Knowledge StudioPreviewRelease notes, 4 November
Context setsPreviewContext sets docs
On by default in GeminiPreview4 November blog post
Launch stages as Google lists them in its posts and docs.

Cost and limits

During the preview, running enrichment tasks and staging unapproved change requests are free, the context sets docs say. Standard BigQuery query processing and Cloud Storage charges apply when a task uses customer-managed resources such as storage buckets or BigQuery agent-analytics tables. Standard Knowledge Catalog metadata storage charges apply once entries are created in a context set.

Knowledge Studio’s docs list what the preview doesn’t cover: Assured Workloads compliance regimes including FedRAMP, CJIS, IL4/IL5 and HIPAA, data residency commitments, customer-managed encryption keys and VPC Service Controls. Google tracks each change from preview to general availability in the Knowledge Catalog release notes.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.