RAG document permissions: identity, retrieval filters and source ACLs
Authenticate the caller, then carry source permissions into chunks and queries. Separate index isolation, retrieval filtering and permission updates, with explicit behavior when authorization is unavailable.
By
Tenten AI Research Team
AI Infrastructure
Published
Updated
Read time
4 min

RAG turns source documents into searchable chunks. Follow that data path when designing document permissions: who made the request, which chunks are eligible, and when the system stops using old permissions after a source change. Successful ingestion and correct answers cannot, by themselves, establish that access controls work.
Establish a trusted identity and authorization decision
The backend should obtain user and tenant identity from an authenticated session, then resolve eligible groups or resources through a trusted authorization service. A department name, group ID or model-inferred role supplied by a visitor cannot directly grant access. Check every protected entry point, including search, source retrieval and download. Deny the requested access when no valid authorization decision is available. OWASP authorization guidance
Azure AI Search security filters illustrate this distinction: they compare identifier strings in a query and do not authenticate those strings. The application must construct trustworthy conditions and prevent callers from bypassing its protected query endpoint. Microsoft security filter documentation
Choose isolation and query conditions separately
Index isolation, retrieval filters and source authorization queries can work together. Map the data and request paths before assigning each layer its responsibility.
| Decision | Question to answer | Constraint to verify |
|---|---|---|
| Tenant or resource isolation | Which data shares an index or namespace? | Can routing select the wrong tenant? How are shared documents updated? |
| Query-time filtering | Which authorized chunks can enter the candidate results? | Does every search entry point apply the complete condition? |
| Source authorization | Use a permission replica or ask the source at query time? | Does the source support the required semantics? What happens on timeout? |
| Checks before model input | Are retrieved content, titles and citations still eligible? | Can reranking, summaries or caches bypass the check? |
Filter timing can also affect result quality. Azure AI Search distinguishes vector prefiltering from postfiltering: filtering after candidate selection can miss eligible results, while prefiltering can increase query work. These are implementation-specific trade-offs. Compare the actual index, filter selectivity and query set before treating a mode as a performance guarantee across databases. Microsoft vector filter documentation
Treat source permission updates as separate work
Chunks need traceable source identifiers and applicable permissions, with access evaluated during retrieval. Do not assume a chunking tool preserves ACLs automatically or ask a prompt to decide which secrets may be disclosed. Citation titles, summaries and shared answer caches can also reveal restricted content. OWASP RAG security guidance
Checking permissions at query time does not establish immediate revocation. Indexes, replicas, identity tokens, group resolution and caches have their own update behavior. The SharePoint indexer preview documentation reviewed on October 12, 2026, for example, describes API-dependent ACL updates and additional refresh requirements for parent permissions. Verify the version and source features used in the proposed deployment. Microsoft SharePoint ACL limitations
Deliver evidence that can be rerun
Put identity sources, authorization services, index scope, update events and failure behavior on the same request-path diagram. Assign an owner to each connection and record versions and test results. Recheck these decisions when changing a connector, database or query mode.
Use the RAG permissions acceptance checklist to test revocation, caches and citations. For a discussion of actual sources and integration scope, start with the RAG knowledge service.
Summarize this page with AI
Related posts

Start with one workflow.
Agree on the next step.
Leave your contact details and question type. We will follow up by email to understand the work and discuss scope.