1min.

1MIN · AI & AGENTS

Apple trains retrieval routers for the queries users actually ask

Apple researchers train small neural networks to improve vector retrieval for a known database and query distribution. SupportNet chooses promising groups by estimating their best possible match, while KeyNet transforms queries before an existing index searches. In the reported Natural Questions experiment, a small SupportNet reached 95% routing recall about 1.9 times faster than centroid routing.

Apple Machine Learning Research · Nov 5, 2026

Why it matters

Retrieval systems can spend computation more effectively when they learn which queries they repeatedly serve.

More context

The explainer covers ICML 2026 research first published July 2. Its benchmark uses roughly 2.6 million passages and 128 clusters; reported latency includes the neural router. Evaluate on representative queries, including distribution changes, before replacing an existing retrieval path.

Keep in mind

Research results assume an upfront training cost and a fixed key database. A routing-recall improvement is not proof of better generated answers or the same speedup on another corpus.

Original read

More engineering ideas, one minute at a time.

Explore more stories