Apple researchers propose learned retrieval method to speed up RAG document searchMachine translation

In a paper presented at ICML 2026, Apple researchers train small neural networks for a fixed document database using examples from a known query distribution to find matches faster. In a test setup with roughly 2.6 million passages, SupportNet reached 95% routing recall about 1.9 times faster than centroid routing; training requires computing the best match for sampled queries in advance.