The One Trace Field That Catches Embedding-Model Mismatches

Embedding-model mismatches between index time and query time are dangerous specifically because they fail silently: retrieval still returns a top-k, generation still produces a fluent answer, and nothing throws an exception. You want to add one piece of per-request trace metadata that would let an on-call engineer diagnose this specific bug class in minutes from traces alone, without needing to reproduce it live. What should you log?

A. The embedding model id and version used for both the index's vectors and the current query's embedding, on every retrieval trace. B. The full text of every retrieved chunk, so a human can eyeball relevance after the fact. C. The wall-clock latency of each pipeline stage, so slow stages stand out. D. The final citation ids the LLM included in its answer.