Enterprise AI Search and Summary Test Record

Enterprise AI Search and Summary Test Record

Sealed source set

Build the answer key before using the AI feature. Freeze document versions, authority order, user roles, permissions, and expected behavior.

The set needs direct-answer, multi-source, stale, conflicting, ambiguous, missing-answer, restricted-source, adversarial-document, and unauthorized-action cases.

Separate answer and evidence scores

An answer can sound correct while citing an irrelevant source. A citation can be accurate while the answer omits material conflict.

Score answer correctness, completeness, qualification, freshness, citation support, citation coverage, source authority, access behavior, abstention, and action authority separately.

Permission cases

Use authorized test accounts representing different roles. A restricted answer should remain unavailable even when the prompt mentions its contents.

Also test overshared source material. Respecting existing permissions does not repair excessive access.

Prompt injection and action boundary

Place safe test instructions inside a source document and verify that the search feature treats them as content rather than system authority.

An answer surface should not invent permission to send, edit, delete, approve, assign, or publish.

Retest triggers

Retest after model, retrieval, citation, connector, corpus, permission, plan, admin, or retention changes. Passing a small set does not prove broad reliability.