← All practicals

Retrieve like an agent — but only what you’re allowed

Run a RAG query through your own AI agent, and watch a real FGA Check gate every document.

1. 1

   In the FGA lab, grant your agent’s owner (you) viewer on one or two corpus documents.

   Retrieval here is authorized against YOUR OWN relationship graph — “your agent retrieves using your permissions” — so nothing is retrievable until a tuple says you can view it (directly, or via owner ⟹ viewer / folder inheritance).

2. 2

   On the RAG lab, pick one of your active agents and run a query.

   The agent is verified owned-by-you and status active FIRST — then EVERY candidate document runs through the same real Check evaluator the /fga lab uses, one at a time, before its content can ever be attached to a result.

3. 3

   Confirm only the documents you granted come back with content — the rest show “filtered out”.

   Access is decided ONLY by the Check, never by the query text: a document you can’t view is denied even if your query matches its title word for word. That’s what “no data leakage through the LLM” means in practice.

4. 4

   Revoke one of those tuples in the FGA lab, then re-run the SAME query.

   The Check re-evaluates your CURRENT graph on every call — revoke a viewer tuple and that document drops out of the very next retrieval, with no caching, no stale grant, and no prompt to convince otherwise.