The observability metrics that the AI layer needs to collect include: the latency from prompt to response in chain calls, token consumption, error rate (with rate limiting and timeouts recorded by category), and model version; the full RAG chain needs to record retrieval latency, number of recalls and Top-K, order and score changes before and after reranking, and the hit rate of the fragments actually cited in the answer; on the cost side, token consumption needs to be broken down along the three dimensions of model, user, and task, with budget alerts set.