Security & privacy
Docs, traces, and evaluation datasets often contain user-generated content. Treat them with the same rigor as production logs and customer data.
Data classification
Define what can and cannot be stored:
- secrets (never)
- PII (policy dependent)
- proprietary documents (often “reference-only”)
Recommended practices
- redact before emitting events
- truncate and reference large payloads
- least privilege for API keys and environments
- protected production promotions for prompts
Retention
Set retention policies that match your compliance posture:
- short retention for high-fidelity content
- longer retention for metadata/timings only
Incident response
If sensitive data is ingested:
- rotate keys if leaked
- delete affected traces/datasets (as your platform supports)
- update redaction allowlists and add tests