I’m relatively new to self-hosting local AI, so maybe I’m missing something obvious. But I’ve been reading through the NeMo Guardrails docs, and the logging default caught my eye.

If I’m running a guardrail layer that logs every single prompt and every response to disk or to a remote endpoint, isn’t that basically creating a complete transcript of every user interaction? In a high-privacy context — like processing internal health notes or financial queries — that log becomes a huge liability.

One compromise I’ve seen is setting `log_all_interactions: false` in the config, but then you lose visibility into what the guardrail is actually blocking or letting through. And if you can’t audit bypass attempts, are you really secure? Feels like a choose-your-own-adventure between privacy and visibility.

Has anyone here deployed NeMo with a “log metadata only, not content” approach? Or maybe a quick hashing of inputs before logging? I’d love to hear what tradeoffs actually look like in practice — not just the theory.