Vectors from two different embedding models are not comparable, even at equal dimensions

Embedding spaces are model-specific. Two models that both output vectors of the same length place concepts in unrelated coordinate systems, so a similarity computed between one model's query vector and another model's document vector is noise that happens to fall in the right numeric range.

This is the failure mode of an incremental model upgrade. Re-embedding only new documents while leaving the old ones alone produces an index where results are dominated by whichever subset matches the query encoder, and nothing errors.

Store the model identifier and the dimension alongside every vector, and make the query path assert they match before searching. Upgrading means re-embedding the whole corpus, so plan for a shadow index built in the background and an atomic switch rather than an in-place migration. The same rule applies to a change in the input template, such as adding or removing a prefix, since that also moves vectors within the same model.