01 /Tag — Access Control
20 articles in this topic.
Apply document permissions before protected passages reach the model. Asking the assistant to hide restricted material after retrieval leaves an unnecessary disclosure and influence boundary inside the answer-generation context.
An answer audit should show which documents actually supported an employee’s conclusion and whether that employee could access them. A list of search hits is not the same as evidence used in generation.
Permission tests should be a release gate for a knowledge assistant, not a one-time project checklist. Search, caching, ingestion, identity, and source-preview changes can each alter what protected information reaches a user.
Check access when a user opens a source citation, even if retrieval was permitted earlier. Roles, sharing links, document status, and account context can change between answer generation and source access.
A knowledge-platform administrator may need to manage indexes without reading every document. Separate configuration, content access, permission administration, and emergency recovery according to actual responsibilities.
Contractor access should match the work they are authorised to perform, including a defined end condition. Avoid granting the full employee knowledge corpus merely because the contractor needs one project collection.
Employee access changes must reach search indexes, caches, ongoing tasks, and source links. Updating one group list does not establish that previously assembled AI context stopped being usable everywhere.
Group membership can arrive late at a search system even after the identity source changes. Measure and design for that delay instead of describing copied group metadata as current permission by default.
A shared document link should not automatically grant the knowledge assistant access. Resolve the source system’s actual sharing semantics and the current user’s authority before retrieving or previewing its contents.
External client documents need a trusted client boundary in internal search. Employees serving several clients should not receive cross-client evidence merely because the documents discuss similar products or problems.
Map company document access before connecting an AI assistant to the content. The project needs to know who can read which resources, how that authority changes, and where derived copies could bypass it.
Search suggestions can leak restricted document titles before retrieval begins. Apply the appropriate permission boundary to autocomplete, recent-query lists, counts, and previews as well as full search results.
Retrieval logs can disclose sensitive information through titles, group names, queries, and source relationships even when passage text is excluded. Design metadata collection around defined diagnostic questions.
Revoked access must affect cached evidence and answers, not just fresh search results. Track which resources a cache entry depends on and re-evaluate whether the current user may receive it.
Row-level and document-level access solve different resource problems. Choose the boundary that matches the underlying data, then preserve it when records become passages, indexes, previews, or generated answers.
Permission to read a source does not automatically justify every answer generated from it. The task, recipient, and disclosure policy can impose an additional boundary on how authorised information is used.
Separate support administration from customer accounts with scoped diagnostic access, explicit repair permissions, time limits, and traceable customer-impacting actions.
Department-based search tests should verify permitted retrieval and forbidden disclosure. Use fictional employees with overlapping roles so the evaluation catches both missing access and accidental cross-team exposure.
A team transfer is both an access-removal and access-grant event. Test the transition rather than only the final group list, because cached context and ongoing tasks may retain the employee’s previous scope.
Re-indexing must preserve source permissions throughout the transition. A new index that contains accurate text but incomplete access metadata can expose documents even when the previous version was correctly scoped.