Question

Your product-search service embeds the query with model M and does ANN retrieval against a vector index of item embeddings. Overnight you upgraded the embedding model from M to M2 and rebuilt the item index with M2. This morning relevance is terrible: searches return near-random items, recall@10 cratered, though the service is fast and returns 200s. Dashboards: the query-embedding service was NOT redeployed and still encodes queries with M, while the item index now holds M2 vectors; the index version tag reads 'M2' but the query path's model version reads 'M'. No latency or error anomaly. How do you triage and fix?