PII redaction

PII redaction strips personally identifiable information that a step doesn't need. Detected values are replaced with placeholders such as [EMAIL_1] or [PHONE_2], and sometimes mapped back afterward, so the final output can contain the real values without the model ever seeing them.

It's one of the most practical privacy controls you can add. It reduces what your provider, logging tools, and tracing platforms receive, limits the damage of a leak, and makes data-handling questions in security reviews easier to answer. It matters most for logs, traces, eval datasets, analytics jobs, and any feature that sends customer content to third-party services.

Redaction is imperfect. Pattern matching catches structured values like emails, card numbers, and phone numbers well, and names, addresses, and free-form identifiers poorly; entity-recognition models do better but still miss things. Over-redaction can break the feature, too: a summary where every name became a placeholder may be useless to the person reading it. Decide per feature what the model truly needs, test redaction on real samples, and never describe it to customers as a guarantee.