Google made LookML indexing in its Knowledge Catalog generally available for Looker Core on Google Cloud on October 26. The release gives agents access to metric formulas, join relationships and mandatory filters alongside catalog context. That makes the business decisions encoded in those definitions relevant to more consumers.
In Google's examples, an agent querying raw tables can omit a completed-order filter or join invoices to line items in a way that multiplies rows and inflates totals. Indexing LookML makes the modeled rules available when the agent generates SQL. Whether a particular answer applies them correctly still needs verification.
What changes with the tool
Google describes its earlier catalog integrations as capturing shallow, file-level pointers and basic dashboard inventories. The new indexing exposes more of the existing LookML model. The semantic logic remains in LookML; the catalog makes it discoverable alongside other enterprise context.
The dbt transition illustrates a different change. In April 2023, dbt Labs announced that it would deprecate support for dbt_metrics with dbt-core v1.6, in favour of MetricFlow. The stated limitations were technical: Jinja macros restricted flexibility, and MetricFlow offered more capable join handling. Users adopting the replacement would need to move their definitions to a new specification, with migration scripts promised to help.
Shreya Shankar's discussion of Data Agent Bench helps explain why that work matters. She reports that providing a semantic layer significantly increased accuracy, and that building the layer was the hard part for most queries. She also stresses the importance of humans understanding what is in the layer.
Keep accountability with the definition
Google's release raises the stakes in one specific way. Once indexing is enabled, Google says synchronization is always on, and changes pushed to production LookML appear in the catalog in near-real time. A change to a revenue measure can therefore reach agents using that indexed logic as well as BI reporting.
Before relying on that wider use, establish who approves a change to total_revenue and where the decision is recorded. A code review should verify that the calculation implements the approved meaning; the accountable business role should approve a change to that meaning.
Write the definition down with a named owner, and keep its approval history associated with the executable metric. During a tool migration, include those records in the work and have the owner confirm that the new implementation preserves the intended calculation. When adding an agent consumer, check its answers against the same approved definition.
Sources
- Unifying enterprise context and semantics: LookML in Knowledge Catalog is now GA, Google Cloud
- Can data agents do real analysis?, Analytics Engineering Roundup
- Deprecating dbt metrics, dbt Labs