Retrieval Authorization

Explanation

Retrieval Authorization checks that, in a Retrieval-Augmented Generation (RAG) system, every document chunk the model fetches and uses to answer a query is one the requesting user is actually permitted to see. This matters because a chatbot answering a customer or a relationship manager might otherwise surface another customer's statements, internal credit notes, or restricted policy documents simply because they sit in the same vector store. It is measured by Retrieval Access Enforcement Effectiveness — the proportion of retrieval events where access controls were correctly applied, validated by replaying queries under different user entitlements and confirming that filtered, unauthorised chunks never reach the prompt. To implement it, attach identity and entitlement metadata (role, branch, customer-ownership, classification) to every indexed chunk, enforce row/document-level filtering at retrieval time before context assembly, and log each retrieval with the user, the candidate chunks, and the allow/deny decision so the score can be computed and audited. The threshold is a target of at least 0.90; a score below 0.85 should trigger investigation, and below 0.80 should trigger remediation such as tightening filters, re-indexing with correct labels, or blocking the affected retrieval path until controls are restored.