RAG Observability Minimum Log Schema per Request Never log raw user PII. Hash user IDs, and redact the query if your policy requires it (see ). LangSmith The retriever run type surfaces chunk scores in the UI. Attach feedback with from your app. Langfuse (Self-Hostable) Langfuse's strengths: self-hosted, prompt versioning, dataset + experiment runner, RBAC. Arize Phoenix (Local, OTel-native) Phoenix exports OpenTelemetry spans natively; point it at any OTLP backend (Jaeger, Tempo, Honeycomb) in production. Comet Opik Opik includes built-in LLM-judge metrics and integrates with CI for regressi…