Skip to main content
Retrieval-augmented generation pipelines retrieve documents from a data set and hand them to the model as context. If the data set contains anything the asking user shouldn’t see, the model will happily summarize it back to them. The fix is to filter retrieval results by the user’s permissions before the model ever sees them.

​
The pattern

  1. The user asks the agent a question.
  2. The retriever pulls candidate documents (vector search, keyword, hybrid).
  3. OpenFGA filters the candidates: for each candidate doc:X, check whether user:Y has can_view. Or call list-objects once to get the full set of documents this user can read, then intersect.
  4. Only the surviving documents are passed to the model as context.
The same model and prompt now produce different — and correct — answers per user, because the context they see is scoped to what they’re allowed to see.

​
Simplified model

Folders nest arbitrarily deep — a viewer on a top-level folder inherits can_view on every descendant folder and document. Before retrieval reaches the model, call list-objects(user:alice, can_view, document) to get every document Alice can read, then intersect with the retriever’s candidate set.

​
Why list-objects matters here

For small set of documents, per-document checks are fine. For larger ones, list-objects is dramatically cheaper: one call returns the full set of documents the user can read, and you intersect that with the retriever’s candidates. This is exactly the case OpenFGA’s reverse queries are designed for.

​
Conditions and contextual data

If access depends on document attributes (classification, region, time window) as well as relationships, use conditions and contextual tuples. Both are evaluated at check time without round-tripping back to your application.
Last modified on September 28, 2026