Which observability record is most useful and safest for debugging a slow RAG answer?

Which observability record is most useful and safest for debugging a slow RAG answer?

1. A correlated trace with timed retrieval and generation spans, token and cost metadata, stable versions, and redacted approved fields
2. A single total-duration number with no model, prompt, retriever, or request identifiers
3. Complete raw prompts, documents, credentials, and user records stored indefinitely