Oracle turns NL2SQL metadata freshness into an operating schedule

Oracle has added three controls that matter more in production than another prompt example: model selection, recurring semantic-store enrichment and asynchronous SQL generation. The release turns metadata freshness and long-running requests into explicit operating choices rather than implementation details hidden behind a chat interface.

Put enrichment on a clock

Oracle’s release note says administrators can choose the Generative AI model used to enrich a semantic store. API callers can also select a model for an individual SQL-generation request; when they do not, the service uses openai.gpt-oss-120b by default. Eligible choices are the on-demand models available in the region and accessible to the tenancy.

That separation gives teams two tuning surfaces: one model can prepare semantic context while another handles a particular question. It also means model changes should be treated as configuration changes and evaluated against known questions, not silently rolled into the service.

The same release adds scheduled enrichment using ISO 8601 durations, with a minimum interval of six hours. An API-based delta refresh updates only database objects changed since the previous enrichment. For warehouses whose schemas, comments or relationships change during the day, this creates a concrete freshness policy: decide how stale semantic context may become, then monitor the enrichment job rather than hoping a user reports a bad join.

Background does not mean faster

SQL generation can now run as a background job through Oracle’s API, SDKs and CLI, but not through the Console. Clients receive a job to monitor and retrieve the SQL after success. Oracle positions this for requests that may exceed an ordinary client timeout; it does not promise lower latency.

That distinction is reinforced in Oracle’s implementation guide, which says background processing prevents a gateway timeout from being mistaken for failure and lets a client expose progress or cancellation. It does not make the database query execute faster.

Keep preparation separate from execution

The guide also describes two Database Tools connections: a higher-privileged connection for reading schema artifacts during enrichment and a lower-privileged connection for end-user queries. Database Tools MCP Server authorizes and executes generated SQL under the configured connection or propagated user identity; NL2SQL itself generates SQL but does not override database permissions.

That is the deployment pattern worth copying: schedule semantic preparation, test model changes, make long work observable, and keep the execution identity narrower than the enrichment identity. Oracle’s current published path generates Oracle SQL against Autonomous AI Database. The company says SQLite and MySQL support are planned, but timing is not committed.