Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Knowledge graphs

Google turns its data catalog into a context engine for agents

At Google Cloud Next in Las Vegas this week, Google announced that it is evolving Dataplex, its data catalog, into what it calls the Knowledge Catalog: “a dynamic, always-on Knowledge Catalog that serves as the universal context engine for your enterprise, helping agents execute complex tasks with accuracy.” It is part of a broader set of announcements Google calls the Agentic Data Cloud.

Google’s framing of the problem is blunt: “When agents lack business semantics and data relationships, this triggers hallucinations, high latency, and stale insights.”

/ 01What Google announced

Google describes three pillars. Aggregation harvests metadata from Google’s own databases and BI tools — BigQuery, AlloyDB, Spanner, Cloud SQL, and in preview Firestore and Looker — plus partner catalogs such as Atlan and Collibra, with preview connectivity to enterprise applications including Palantir, Salesforce Data360, SAP, ServiceNow and Workday. Enrichment uses Gemini to extract metadata from unstructured and multimodal content and to generate descriptions and business glossaries automatically, with verified queries and “semantic guardrails” in preview. Search offers semantic search, now generally available, that respects the access permissions of the source systems.

SiliconANGLE describes the catalog as a “map of business meaning” that scans documents including PDFs, presentations and images, extracts entities, studies the relationships between them, and builds a navigable schema for agents. IT Brief reports that it enriches context over time using activity data, and that Google’s Deep Research Agent in Gemini Enterprise, in preview, already runs on it. Google names Bloomberg Media as a customer using it to power a data access agent.

The blog post closes with its thesis: “Stop forcing your agents to guess the unwritten rules of your business.”

/ 02Why it matters

For several years the default enterprise pattern for grounding AI has been retrieval: embed documents, find similar passages, paste them into a prompt. Google’s announcement is an admission, from one of the largest AI vendors, that similarity search over raw content is not enough for agents that must act. Agents need to know what things are, how they relate, and which definitions the business actually uses. In other words, they need a semantic layer.

That is a significant shift in where the industry places value. A catalog used to be an inventory of tables for data engineers. Google is repositioning it as the knowledge an agent reasons with. Even the language — extracting entities, mapping relationships, maintaining a map of business meaning — is the language of knowledge graphs and ontologies, though Google’s post does not use either term.

/ 03Context, ontology and provenance

The interesting questions are in the details. A context layer generated largely by a model, from metadata and documents, will reflect what the model inferred. That is a good starting point and a risky end point. The business glossary that defines “active customer” or “readiness” is exactly where an organisation needs its own experts to have the last word, and where it needs to know who decided and when.

Two properties separate a context layer an agent can lean on from one that just moves guessing upstream. The first is a formal model: classes, relationships and constraints that are explicit and consistent, so that two agents asking the same question resolve it the same way. Upper ontologies such as BFO exist precisely to make those models line up across domains. The second is provenance: every entity and relationship should point back to the source that asserted it, so an agent’s answer can be traced through the context it used to the data underneath.

Location matters too. Google’s catalog naturally centres on Google Cloud, with partner connectors. Organisations in defense, intelligence, health care and finance often need the same capability over data that cannot leave their own environment.

When the largest vendors say agents need business semantics, the debate shifts from whether to build a semantic layer to who controls it.

/ 04What to watch or do

  • Preview to GA. Much of the enrichment and the enterprise connectors are in preview. Watch what reaches general availability and on what terms.
  • Human curation. Look at how generated glossaries and relationships are reviewed and approved. Automated context is only as good as the review process behind it.
  • Portability. Ask whether the context layer can be exported in open formats such as RDF and OWL, or whether it lives only inside one platform.
  • Evaluation. Google describes a framework for measuring context quality. Apply the same discipline to whatever semantic layer you build.
Aktiver’s take — Enterprise data

Google is conceding that retrieval alone leaves agents guessing. The open question is whether a context layer drafted by a model can carry the authority of definitions a business has to defend.

In an enterprise graph, a term like active customer is a class with a formal definition, an owner who approved it, and links to the tables and documents that instantiate it. Each relationship records which source asserted it and whether an expert has reviewed it. An agent answering a revenue question can then show which definition it applied and who signed off on that definition. Ingestion Hub produces this kind of ontology from PDFs, SQL, CSV and JSON, with subject-matter experts refining the result and BFO alignment keeping separate domains compatible. Export to OWL or RDF keeps the model portable across platforms, and on-premises or hybrid deployment covers data that cannot move to a public cloud.

How we’d model it

Business termFormal definitionData ownerTableRelationshipSource systemdefinedByapprovedByinstantiatedInassertedByBusiness termFormaldefinitionData ownerTableRelationshipSource systemdefinedByapprovedByinstantiatedInassertedBy
  • Business term —definedBy→ Formal definition
  • Formal definition —approvedBy→ Data owner
  • Business term —instantiatedIn→ Table
  • Relationship —assertedBy→ Source system

See where this applies →

Sources — read the originals
  1. Introducing the Google Cloud Knowledge Catalog — Google Cloud Blog cloud.google.com ↗
  2. Google delivers connective tissue for autonomous AI agents to access data without restrictions — SiliconANGLE siliconangle.com ↗
  3. Google Cloud launches Agentic Data Cloud for AI agents — IT Brief Australia itbrief.com.au ↗
Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading