Oracle has released Trusted Answer Search 26.3, an update to its product that turns natural-language questions into governed answers inside applications. Oracle announced it on its Database blog on November 12, 2026. The release focuses on measuring search quality, reviewing weak queries and cutting latency.
For teams that already run it, the new tools give a clearer picture of what users search for and where results fall short. For everyone else, it is a good moment to learn what the product does and where it fits next to vector search.
What Trusted Answer Search Does
The product maps a user's question onto a curated list of targets that your own team defines. A target can be a URL, a SQL statement, a script or an MCP endpoint, and it can take parameters. The service also works out the input values a target needs from the wording of the question.
It then hands the chosen targets back to the calling application. The application stays in charge of permissions and of carrying out the action, such as sending the user to a page or calling an endpoint. That keeps the answer governed rather than generated freely.
Oracle also points to agents. As a company's tool catalog grows, stuffing every tool description into an agent's prompt adds cost and makes the right pick harder. An agent can instead search a curated MCP tool catalog and find the relevant tools when it needs them.
What Is New in 26.3
- Search analytics and telemetry: administrators can review how often results get clicked, which searches get no click at all, which ones users flagged as failures, plus latency summaries and deeper performance detail.
- Evaluation sets: teams can collect queries into evaluation sets and use them to test a planned change first.
- Query review: search-space administrators and experts can flag queries for follow-up, add review notes and track how each one was resolved.
- Search-space cloning: an entire search space can be duplicated to seed a different application or scenario.
- Shared embedding model: sessions in the database can reuse a single ONNX embedding model, which avoids reloading it each time and lowers latency.
- Regular expression routing: administrators can define patterns that send matching queries straight to a chosen target.
An Example From Oracle
Oracle shows a fictional brokerage interface. When a user types "my equity allocation", the first hit is a page named Allocation Explorer and the asset class value is filled in as equities. Two more pages are listed after it, and for both the account scope is filled in as all accounts.
Oracle also says its own internal bug-reporting website uses the product in production. In an internal test, the precision at the first, third and fifth results was 94%, 97% and 100%. Oracle adds that these figures come from one internal evaluation and may not match other deployments.
What It Means for Developers
Search that guesses is hard to trust, and search that cannot be measured is hard to improve. The analytics and review tools address the second problem. Searches with no clicks and user-reported failures show you which targets or sample questions need better descriptions.
Regular expression routing is useful for predictable inputs such as order numbers or ticket codes. Instead of hoping an embedding model matches the pattern, you can send it to the right target with a rule. Cloning helps when several apps share a similar catalog.
If you build with vectors in Oracle Database, this is a higher-level option for navigation and action search. Pair it with your own evaluation set so you can see whether a change to a description helps or hurts. For related news on agents in the Oracle stack, read our report on LangChain Deep Agents on Oracle AI Database 26ai.
How to Get Started
Oracle links a download for Trusted Answer Search 26.3, a product page and documentation from its announcement. Start with a small search space, define a handful of targets and add sample questions for each. Then use the new analytics to see how real users behave.
Conclusion
Release 26.3 makes Trusted Answer Search easier to run and tune. If you use it, turn on the analytics first and review the queries that get no clicks. If you are new to it, read Oracle's documentation and try it on one application.
Sources
See my articles more often when you search Google.
