Apple’s research accelerates approximate maximum inner-product search using KeyNet to transform queries and SupportNet to predict cluster-level scores. In a Natural Questions experiment with roughly 2.6 million passages and 128 clusters, a large KeyNet reduced latency at 99% recall from 1.8 ms to 0.23 ms, including query transformation. The released code requires computing cluster-level targets and offline training. These are retrieval-stage results with fixed key data and representative queries, not evidence of equivalent improvements in end-to-end RAG speed or generated-answer accuracy.
Back to Tech
Apple improves vector retrieval with learned query transformations
Offline training on a fixed corpus has a cost, and faster retrieval does not guarantee better generated answers.