Embedding Drift

Embeddings are only comparable when produced by the same model and version. When a provider updates an embedding model, or an organization switches providers, vectors created earlier occupy a slightly different space from vectors created after — similarity scores across the boundary become unreliable, and retrieval quality degrades in ways that are hard to attribute because nothing fails. The symptom is gradual: results that used to rank first drift downward, recall for older material declines, and the cause is invisible without deliberate monitoring. Remedies are re-embedding the affected corpus or pinning the embedding model, and the choice is largely about how much content is involved.

Because Centralpoint indexes at the record level and holds the index in the organization's own environment, re-embedding is a controllable operation rather than a vendor event. Model selection happens at runtime, so an embedding change is a decision the organization schedules rather than one that arrives unannounced.