An internal knowledge assistant should not become a new route around the permissions that already protect business information. A user who cannot open a document should not be able to obtain its contents by asking the assistant to summarise it, cite it or combine it with other sources.
Permission-aware retrieval means carrying access rules through the entire answer path. The relevant controls begin before search results reach the model and continue through citations, conversation history and caches. A reassuring instruction in the prompt is not a substitute for enforcing those controls in the application.
Start with the source permissions
Map the repositories the assistant will search and identify how access is represented in each one. Some systems use groups, others use individual shares, and some contain inherited permissions with local exceptions. The implementation must understand the actual rules rather than assume that all documents in one folder have the same audience.
Decide which identity the search operation uses. A broad service account may be convenient for indexing, but its ability to read a document does not establish that the requesting user is allowed to receive it. The query path needs a reliable way to apply the user's current entitlement.
Keep the first scope manageable. A single well-understood repository is a better starting point than combining several systems with incompatible access models and hoping that a universal prompt will compensate for the differences.
Carry permission metadata with indexed content
When documents are divided into searchable passages, retain the metadata needed to enforce their access rules. A passage without its document identity and permission context can become an orphaned copy of restricted information. The same concern applies to extracted tables, captions and attachments.
Record how permission changes reach the index. Revoking access in the source system is not enough if the search layer continues using an old entitlement snapshot. Define the expected update delay and the behaviour during an uncertain or failed synchronisation.
Where access cannot be established confidently, choose a conservative result. The assistant can explain that it cannot retrieve the required evidence or route the user to an approved process. It should not fall back to a broader search identity simply to avoid an incomplete answer.
Enforce access before generation
Filter retrieval results using the authenticated user's permissions before sending passages to the model. Once restricted content is included in the model context, asking the model not to reveal it becomes a weaker boundary. The application should avoid providing information the task is not authorised to use.
Apply the same discipline to tools that fetch additional documents. An initially permitted search can lead to a later tool call for a related source. That follow-up must be checked independently; permission to view one document does not automatically extend to its references or attachments.
Do not let generated tool arguments select an unrestricted identity or override access filters. Identity and permission scope belong to trusted application logic. The model can propose what to search for, but it should not decide whose authority the search runs under.
Treat citations as part of the access boundary
A citation can disclose information even when the answer text is carefully limited. Document titles, folder names and snippets may reveal the existence of a confidential project. Check what metadata is safe to show, not only whether the full document body is protected.
Make citation links respect the source system's access checks. Avoid creating public copies of restricted documents to make citations convenient. If an authorised user cannot open a link because of an integration fault, fix the access path rather than bypassing it through a downloadable export.
Test how the assistant explains missing evidence. It should not identify restricted sources unnecessarily while saying that access was denied. A neutral explanation that the available authorised material does not support an answer may be more appropriate than naming the inaccessible document.
Review conversation and cache behaviour
Cached answers can preserve information after the context that authorised them changes. A shared cache must not return one user's answer to another user with different permissions. Cache keys and invalidation rules need to account for the relevant identity and access scope, or the response should remain uncached.
Conversation history creates a related challenge. A user may lose access to a repository after an earlier session retrieved material from it. Recheck current permissions before reusing retrieved passages or derived summaries in a new answer; exclude material whose access has been revoked or cannot be verified. Otherwise, history can bypass an otherwise correctly filtered search. Whether a historical transcript is retained or remains viewable is a separate governance decision, not permission to reuse its evidence in new generation.
Avoid treating a role name as a complete permission model when users with the same role can access different records. The cache and session design should reflect the real entitlement structure, including changes and exceptions that matter to the organisation.
Test with paired users and changing access
Create two authorised test accounts with different access to a synthetic document set. Ask the same questions through both accounts and inspect answers, citations, search results and logs. The restricted account should not receive information merely because the other account asked first.
Then change permissions and repeat the test. Include document removal, group membership changes and a failed permission synchronisation. These cases expose weaknesses that a static test with one administrator account will not reveal.
Test indirect questions too. A user might ask for a comparison or aggregate summary rather than request the restricted document directly. The system still needs to ensure that the evidence used to build the answer is authorised, even when the final wording does not quote a source verbatim.
Make operational ownership explicit
Assign responsibility for permission mapping, synchronisation failures and access-related user reports. The team maintaining the model prompt may not own the document repositories, so the incident path needs to cross those boundaries without leaving the issue unresolved.
Keep useful audit evidence without unnecessarily copying restricted content into broadly accessible logs. Record source identifiers, entitlement checks and denial reasons at an appropriate level. Investigators should be able to reconstruct what happened while preserving the same care around access that the product promises users.
A private knowledge assistant is only as private as its complete information path. The desired outcome is not merely an answer that sounds careful. It is an enforceable rule that the assistant uses and reveals only the information the current user is entitled to receive, including when permissions change after the system has gone live.