SharePoint, Google Drive and your CRM know who may read a file. The vector index usually does not. One embedding job turns a restricted folder into public knowledge.
Tenants share a collection
In multi-tenant products, one collection with a tenant column is only safe if every query filters on it. One missed filter in one code path is a cross-customer breach.
Documents carry instructions
Anyone who can get a document into the index can get text into the prompt. Without provenance and scanning, the knowledge base is an injection channel.
Reference architecture
Permissions enforced before the model, citations checked after it
Identity claims flow into the retrieval query; the guard rejects any answer whose citations are not in the retrieved set.
The four controls
What we design and implement
Document-level ACLs
Every chunk carries the permission list of its source document: user ids, group ids or roles, copied at ingestion and refreshed on change. Permissions live in metadata, not in the prompt.
Tenant isolation
One partition, collection or schema per tenant where the database supports it; a mandatory tenant filter injected by the data layer where it does not. The application code cannot forget it because it never writes the filter.
Retrieval filtering
The retriever takes the caller's identity claims and applies tenant and ACL filters inside the vector query, before ranking. Post-filtering after top-k is not acceptable: it leaks through counts and can return empty results for authorised users.
Citation enforcement
The model must cite chunk ids for every claim. A guard verifies each id exists in the retrieved set and that the quoted text appears in it. Uncited or mis-cited answers are rejected and regenerated, or shown with a visible warning.
Comparison
Naive versus secure RAG
Concern
Naive RAG
Secure RAG
Permissions
Lost at ingestion
Stored per chunk, filtered per query
Multi-tenancy
Tenant column, filter optional
Partition or enforced filter
Filter timing
After top-k, if at all
Inside the vector query
Document provenance
Unknown
Source, uploader, time, hash on every chunk
Injection in documents
Pasted straight into the prompt
Scanned, wrapped as data, flagged
Citations
Decorative
Verified against retrieved chunks
Revocation
Stale until re-index
Event-driven removal + reconciliation
Logging
Full prompts, forever
Chunk ids and user, redacted, time-limited
How we work
Design review or full build
Map sources
Where documents come from, who may read them, how permissions are expressed, how often they change.
Choose the isolation model
Partitions, collections or enforced filters, based on tenant count, data volume and your vector database.
Build ingestion and retrieval
ACL-aware ingestion, identity-aware retriever, citation guard, revocation handling.
Test and measure
Cross-tenant and ACL probes in CI, latency before and after filtering, answer quality on your eval set.
A typical design review takes one to two weeks. A full secure RAG build for one source system runs four to eight weeks, including evals. Pair it with AI data privacy work if documents contain personal data.
Why is RAG a security problem? The model only reads documents.
Because retrieval copies documents into an index that usually has no access control, then pastes the best matches into the prompt. If the retriever does not check who is asking, the model will happily summarise a document the user could never open. See our RAG explained guide for the mechanics.
Can we enforce permissions with a prompt instruction instead?
No. Once a chunk is in the context window the model has already seen it, and models do not reliably follow restrictions about what to repeat. Permissions must be enforced in the retrieval query, before the model is called.
Which vector databases support this?
All the ones we work with: PostgreSQL with pgvector, Milvus, Pinecone, Qdrant and Weaviate support metadata filters or partitions that carry tenant and ACL data. The design differs; the principle does not. See vector database selection.
How do you keep permissions in sync when documents change?
Ingestion records the source system's ACL on every chunk and listens for changes. Revocations and deletions remove or re-tag chunks within an agreed window, and a periodic reconciliation job compares the index against the source.
Does this slow retrieval down?
Filtered search costs a little, especially when the filter is very selective. We design partitions and indexes so that tenant and ACL filters are cheap, and we measure latency before and after.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.