Oracle adds search-review tools and direct routing to Trusted Answer Search

Oracle released Trusted Answer Search 26.3 on November 12, 2026, adding tools to inspect unsuccessful searches, review queries and route predictable requests directly to a chosen destination. Available to download, the update also lets teams copy search configurations and share a search model across database sessions to avoid repeatedly loading it.

Finding a destination is not executing an action

Trusted Answer Search maps everyday questions to curated destinations rather than carrying out the requested work itself. Those destinations can be web addresses, database queries, scripts or tool endpoints. It also extracts the inputs a destination needs from the question, then returns the selected target or targets and those inputs to the calling application.

That leaves an important boundary intact: the application must enforce the user’s permissions and perform the action. Returning a web address does not redirect the user by itself; returning a tool endpoint does not call that tool. Search selects the route, while the application controls what happens next.

Oracle illustrates the mechanism with a fictional brokerage interface. A search for “my equity allocation” ranks Allocation Explorer first and supplies EQUITIES as its asset-class input. Two other results, Portfolio Map and Account Overview, receive an all-accounts input. The example shows both parts of the job: choosing relevant destinations and resolving their required inputs.

The same discovery approach can serve AI agents. Oracle says agents can search a curated tool catalog on demand instead of loading every tool description into their working context. The company presents this as a way to keep that context focused as enterprise tool catalogs grow.

Missed searches get a review trail

Version 26.3 gives administrators several views of search outcomes. Analytics cover click-through rates, searches that receive no clicks and searches users report as unsuccessful. Latency summaries and detailed performance breakdowns address the speed of the search experience alongside those measures of user interaction.

Administrators and search-space experts can flag individual queries for follow-up, attach review notes and track their resolution. This creates a record of which questions needed attention and what teams did to improve the results, rather than leaving the review process at a list of unsuccessful searches.

Teams can also add queries to evaluation sets to assess proposed changes. Oracle identifies several things administrators can refine: the targets themselves, their descriptions, sample questions and the sets of values targets accept. The analytics and review tools provide a basis for deciding which of those elements needs work.

Reuse configurations, skip repeated model loading

The release also adds search-space cloning. Teams can copy an entire search space and use its established configuration as the starting point for another application or related use case, instead of building that search experience from scratch.

For search performance, database sessions can now share the same ONNX embedding model used by the search system. Oracle says avoiding repeated model loading reduces search latency. The change concerns reuse across sessions, rather than requiring each session to load that model again.

Administrators can define regular expressions—text-matching rules—to route matching questions directly to a chosen search target. Oracle describes this option for queries containing a known phrase or following a predictable identifier format. It gives administrators an explicit route for those recognizable inputs.

Oracle already uses Trusted Answer Search on its internal bug-reporting website. In testing that implementation, the company measured precision@1 of 94%, precision@3 of 97% and precision@5 of 100%. Those internal evaluation results are not necessarily representative of other deployments, Oracle cautions.