Treat Vector Databases as Caches

A vector database is a derived index, not a durable source of truth. Embedding model upgrades invalidate all stored vectors, causing silent retrieval drift. The raw corpus and reproducible pipeline are the durable assets. Teams should practice rebuilds and store provenance metadata to ensure the index can be regenerated.