Why scrub on the way in and restore on the way out
The moment your RAG pipeline touches customer emails, support tickets, or employee records, you have a compliance surface. Anything sent to a hosted LLM lives in that vendor log retention. A privacy-aware pipeline scrubs PII on the way in and restores the original entities on the way out. The user sees their real names. The provider sees placeholders.
Scrub in, restore out
The anonymizer sits at the boundary. Every LLM call happens on scrubbed text. The final answer is de-anonymized before it leaves the pipeline.
Why reversible? Because answering "what did Alice say about the refund?" with a response that refers to PERSON_1 is useless to the user. The pipeline needs to recover the real names before returning, but only after the LLM has finished generating. The mapping never leaves the server, and the scrubbed versions are the only text the LLM ever sees.
Quiz: Quiz
Loading practice…