Embedding Drift Exposure

Embedding Drift Exposure is the measure of how much a client's retrieval quality decays when the model that generated its vectors is replaced, retrained, or deprecated. Vectors are not portable across embedding models: swap the encoder and every stored vector must be regenerated, which means re-embedding the full corpus, re-indexing, and re-validating answers. Agencies that treat the index as a one-time build inherit this exposure silently, and it surfaces as a client complaint about wrong answers rather than as a line item. The framework asks three questions at scoping: which encoder produced these vectors, who controls its lifecycle, and what the re-embedding cost is at current corpus size. A managed platform such as Pinecone absorbs some of the operational work but not the model dependency; a self-managed engine like Qdrant keeps the vectors in agency hands while leaving the re-embedding run on your delivery team. Budget the regeneration pass into the retainer before the first index ships.