Verified, then logged.
Sun 22 Nov 2026 · 00:19 KST / 15:19 UTC한국어 | ENSearch /

Breaking01 / AI

Apple's KeyNet cuts RAG search time by up to roughly 8 times

The research learns query patterns in advance to cut RAG search costs, and Apple's post does not give the pre-training time.

In a data center, only a few drives in an open server's long row show amber activity lights, while an engineer stands blurred in the background with their back to the camera. AI-generated image

Apple's KeyNet, introduced on 5 Nov 2026, cut search time on fixed databases by up to about 8 times, according to company-published figures in an Apple research post.

The figure is the time taken to reach 99% recall, the share of correct answers found. In a paper presented at ICML 2026, Apple researchers said they used small neural networks for maximum inner product search, a method that finds the stored vector best matching a query. This search is central to retrieval-augmented generation (RAG), which finds documents and attaches them to a question, and the networks are trained once, offline, for each database to speed it up.

Two models. According to Apple, SupportNet predicts the highest score in each cluster, and that score determines where a query is sent. KeyNet predicts the best key, a stored database vector, and feeds the altered query into an inverted file index (IVF), a structure that divides stored vectors into clusters.

The research post confirms that Apple tested the models on about 2.6 million Natural Questions passages. In company-published figures, KeyNet's largest model reached 99% recall about 8 times faster (1.8 ms to 0.23 ms). SupportNet's small model reached 95% routing recall, the share of queries sent to the correct cluster, about 1.9 times faster (3.36 ms to 1.74 ms).

Apple wrote that KeyNet is the choice when inference speed is the bottleneck. It recommended SupportNet when calibrated scores are needed. Systems with a very large number of keys may need both models, it added.

Comparison baseline. Apple explained that standard IVF decides where to send a query by comparing it with centroids, the central points of clusters. When only 5% of the database was scanned, routing recall in company-published figures was 87.5% for centroids and 92.6% for SupportNet M. The post states that the earlier timing figures include the cost of running the neural networks.

Apple's post does not give the pre-training time or say whether the method is used in Apple products.