How to trace a failed RAG answer

Find the broken span before tuning prompts.

1. Reproduce with same request ID

2. Open trace: embed → retrieve → generate

3. Check retrieval count and scores first

4. If 0 chunks → index, filter, or embed model mismatch

5. If chunks OK but bad answer → prompt or model tier

6. Redact PII in stored traces

**Try the lesson:** `llm-tracing-lab` in Lane D