Embedding model swaps aren't drop-in replacements

Every AI team that consolidates its retrieval stack behind a new vendor eventually discovers that the same documents quietly come back in a different order. The cause is not the corpus; it is the math underneath. Vendor-specific similarity geometry is the reusable category: each embedding model is not just a different dictionary of vectors, it is a different ruler for measuring how close two pieces of text are. Treat that ruler as portable, and a "drop-in" migration quietly breaks the threshold the retriever was tuned against.

Rozmus and van der Putten show the shape of the problem. In their paper and a public writeup, they probe the score relationship between Titan and Ada embedding models and find it is non-linear, with different ranges across the score distribution. Titan models of different sizes behave more like each other than Titan and Ada do, even though all three are sold for the same job. The finding is from a single unreplicated preprint (Rozmus and van der Putten, arXiv:2608.05857, Discovery Science 2026); it is the authors' own, not yet independently validated. The naive read says a cheaper model is a swap; the actual pattern says the score scale moves with the vendor.

The mechanism is repeatable. Map the score distributions, not the raw vectors, before you migrate. Build a probe set, not just a benchmark, and recalibrate the threshold on the new model. Even then, the authors flag that the probe set inherits the biases of whatever synthetic-query generator built it. The next migration story worth writing will not be a price comparison; it will be a score-geometry comparison, and the teams that learn to read geometry will keep their retrieval quality while the rest quietly accept whatever reordering the new vendor hands them.