Google Cloud makes LookML indexing generally available

Google Cloud has made LookML semantic indexing in Knowledge Catalog generally available for Looker Core on Google Cloud. The release is aimed at enterprises using AI agents with business data.

It brings LookML business logic into Knowledge Catalog by default, allowing organisations to expose both data catalog metadata and metric definitions to Gemini Enterprise and custom AI agents from the same environment.

The integration is intended to address a common problem as companies move AI agents from experiments into production. Data teams may be able to show agents where data sits, how it is tagged and how it moves across systems, but that does not necessarily tell them how revenue, churn or other business measures should be calculated.

LookML is Looker's modelling language for defining joins, filters and metric formulas. Knowledge Catalog aggregates metadata across structured data and documents, including lineage, governance rules and privacy tags.

Under the new arrangement, resolved LookML models, Explores, Views, lineage and dashboards are indexed directly into Knowledge Catalog. Synchronisation is enabled by default, and updates to production LookML are reflected in Knowledge Catalog in near real time.

This means an AI agent can retrieve not only table descriptions or glossary terms, but also a metric's underlying definition, join path and source lineage in one place. Google argued that this reduces the risk of AI systems generating inconsistent formulas or inflated totals through incorrect joins across one-to-many relationships.

Semantic layer

Google set out the distinction between broad enterprise context and deterministic business semantics. In its view, data catalogues can show what data exists and what governance applies, but they do not by themselves define how a business should calculate figures in SQL.

That gap matters when large language model agents write queries directly against warehouse tables. Without explicit semantic rules, such systems may assume the wrong fiscal calendar, include transactions that should be excluded, or expose restricted data fields if they rely only on raw schemas and general descriptions.

The indexing covers several layers of Looker metadata. At the model level, Knowledge Catalog captures descriptive metadata, connection routing, include hierarchies and validation tests. At the Explore level, it records join logic, relationship cardinality, required joins and mandatory filters. At the View level, it extracts compiled SQL expressions, primary keys, timeframes, synonyms and derived table logic.

Dashboards, Looks and tiles are also catalogued, including filters, chart types, ownership and usage statistics. This allows users to trace lineage from BigQuery tables to individual dashboard tiles.

Agent workflows

Google contrasted three approaches to enterprise data agents. The first is a database agent with no semantic or catalog context, which writes SQL directly against warehouse tables. The second uses Knowledge Catalog for metadata and governance but still relies on the model to infer metric logic. The third combines Knowledge Catalog with LookML indexing, allowing the agent to use both governance information and pre-defined SQL relationships.

In Google's description, the third model gives Gemini Enterprise and similar systems a two-step process. The agent first searches Knowledge Catalog for governance information and the relevant LookML objects, then generates SQL grounded in those indexed definitions rather than inferring calculations at runtime.

Google also drew a distinction between this setup and Looker Conversational Analytics. It described Conversational Analytics as a BI assistant for users working within curated Looker Explores, while Knowledge Catalog combined with Looker is positioned for broader enterprise agent use across BigQuery and other data sources alongside unstructured documents.

The release applies to Looker Core on Google Cloud and integrates with Gemini Enterprise through Knowledge Catalog. Users can attach a BigQuery, Spanner, Cloud SQL or AlloyDB data store in federated query mode, after which Knowledge Catalog support is enabled automatically.

Access controls remain tied to end-user credentials, meaning users can discover and query only the data assets they already have permission to view. The assistant uses those permissions when looking up indexed LookML definitions, join logic and starter queries in Knowledge Catalog.

Google said the result is that agents can move from raw SQL guesswork to using governed business definitions and catalog context from the same system.