Voice AI agents often handle sensitive customer conversations. Teams should not have to choose between protecting that data and having the visibility needed to improve their agents.
LiveKit PII Redaction automatically removes personal information from transcripts and audio recordings before it is stored. It supports 41 types of PII and uses conversational context to detect sensitive information in voice interactions.
PII Redaction is included with LiveKit Agent Observability at no additional cost.
Nice. One thing worth knowing before enabling it: it changes what traces keep on failures.
With redaction on, record_exception keeps the exception type but replaces the message with “exception details redacted” and drops exception.stacktrace, and spans are also created with record_exception=False. Reasonable, since exception text isn’t lk.pii.* tagged so it can’t be selectively stripped, but debugging a prod incident looks quite different. Is a separate non-redacted project for repro the intended pattern here? (on 1.7.0)
Those kind of production incidents are always tricky aren’t they. At that stage I assume the issue was missed during earlier testing so yes, the aim would be to reproduce in a non-redacted project (if the root cause isn’t clear from other indicators, e.g. it aligns with an update to the agent)
