Privacy Policy enforcement – Response Filtering

Explanation

Privacy-policy enforcement via response filtering checks whether sensitive or confidential items that surface inside retrieved context are redacted or filtered before they are used to compose an answer — the retrieval-side guard for RAG systems. It matters immediately because a RAG assistant can pull a document chunk containing another customer's account details into context and echo it back; catching that at retrieval upholds RBI Safety/Resilience, NIST MANAGE 3.1 and ISO Clause 6.1.2. It is measured by the Privacy-Aware Retrieval Rate (PRR) = Sensitive/Confidential Items Redacted or Filtered at Retrieval / Total Sensitive Items Surfaced in Retrieved Context, on a 0-1 scale with target approaching 1.0 — the share of sensitive items appearing in retrieved context that were actually redacted or filtered before reaching the generator. To implement, insert a privacy filter between the retriever and the generator in the RAG pipeline that scans retrieved chunks, redacts or drops sensitive items, and logs each surfaced item, the action taken and any leakage as evidence. The threshold is a PRR of at least 0.95, with at least 0.98 recommended for critical systems; scoring below it means sensitive retrieved content is reaching generation unfiltered, a breach that triggers retriever-filter tuning and re-test, and escalation to the 0.98 standard for critical banking workloads.