Embedding Drift
A shift in the distribution of embedding vectors after model, data-source, or preprocessing changes, potentially invalidating previously calibrated similarity thresholds and retrieval behavior.
A similarity threshold does not necessarily preserve its meaning when the embedding model or preprocessing pipeline changes. If vector norms, cluster density, or inter-class distances move, previously calibrated cosine or L2 cutoffs can produce different false-positive and false-negative rates.
A model update should therefore be validated on a fixed reference set, including distribution statistics, nearest-neighbor ordering, and decision thresholds. Image feature vectors and matching and audio feature vectors and matching show two applications of the same problem.