A BigQuery user report exposes a 10-term ceiling in catalog grounding

A BigQuery practitioner has documented a potentially serious gap between enterprise glossary coverage and the context that reaches a conversational data agent. In a September 3 Google Developer Forum post, the user says a BigQuery Data Agent imported only 10 business terms from Knowledge Catalog even though the catalog contained more than 1,000 terms and hundreds were linked to columns.

This is a user report, not a confirmed Google product limit. But it is unusually useful evidence: the post includes screenshots, describes a successful Dataplex discovery run and distinguishes working terms from missing ones. The 10 imported terms retained their descriptions and business logic. For a linked term outside that set, the author says the context retriever failed to supply the definition; the agent then tried an INFORMATION_SCHEMA.COLUMN_FIELD_PATHS query that was blocked because it fell outside the conversation’s context definition, before falling back to general model inference.

The documentation promises an import path, not its capacity

Google’s data-agent setup guide tells builders that they can review business glossary terms imported from Knowledge Catalog that apply to an agent’s selected knowledge sources. It also directs administrators to edit those global terms in Knowledge Catalog and then return to BigQuery to view the changes. The guide does not state a 10-term ceiling or expose a setting for changing the number imported.

That omission matters because the catalog and the agent serve different jobs. Google’s Knowledge Catalog integration overview describes column-level metadata, business context and a context graph intended to ground agents in approved enterprise semantics. If only a small subset crosses into an agent, a well-governed catalog can still produce an under-grounded answer path.

Treat glossary ingestion as a deployment test

Teams should not assume that linking a term to a column proves the term is available at inference time. Before publishing an agent, inspect its imported-term list and test questions that depend on terms near the beginning, middle and end of the intended set. Record which definition the agent used, not merely whether it generated valid SQL.

For critical metrics, builders can also add agent-local context, explicit instructions or verified queries—the setup guide supports all three—but those are compensating controls, not proof that catalog import is complete. They create another layer to maintain and should be reconciled with the governed glossary.

The immediate unanswered questions are whether the observed 10-item result is a fixed preview limit, a retrieval bug or a configuration-dependent behavior, and how selection is ordered. Until Google answers the forum thread or documents the behavior, practitioners should treat catalog-to-agent transfer as a bounded, testable interface rather than a transparent inheritance mechanism.