The tricky thing is how we're going to handle PII on a lot of different places in the future

The tricky thing is how we're going to handle PII on a lot of different places in the future

If we do PII redaction inside the ChatContext, we can keep changes to other places minimal, right?

The PII isn't realtime, it's going to be done at the end of the session. (On LiveKit Cloud)
So it's already too late, another version of the chat_ctx is already inside some traces (in the attributes)

I see. But currently, we are still reporting LLM output, which might also contain PII. Do we intend to:

1. Keep the codebase as is: no input report, but keep output;
2. Or drop output too;

Either way, this would create a discrepancy between Agent Observability and Otel traces for the moment. If we are going to support Otel PII handling eventually, I think we should at least provide consistency for now, especially considering users have to explicitly set up Otel today.

I like how Logfire handles this by:

1. providing a set of predefined regexes for typical sensitive data
2. providing a callback for more control (but the user takes the responsibility for it being slow/complex)
   , similar to how Langfuse handles sensitive data too.