Skip to content
Research · Nov 7, 2026

Apple researchers propose neural support functions to speed up retrieval for RAG

A new method uses learned support functions to route queries more accurately than centroid-based clustering in maximum inner product search, reducing compute while maintaining accuracy.

Trust79
HypeLow hype

1 source · cross-referenced

ShareXLinkedInEmail
TL;DR
  • Apple ML Research describes a technique to speed up retrieval-augmented generation by replacing centroid-based routing with learned support functions trained on real query distributions.
  • The approach trains small neural networks offline to mimic the support function of a key set, enabling faster and more accurate routing of queries to relevant clusters.
  • Evaluated on ICML 2026 benchmarks, the method reduces retrieval latency while maintaining or improving accuracy compared to exhaustive and approximate baselines.

Apple Machine Learning Research describes a method to accelerate retrieval in retrieval-augmented generation (RAG) systems by replacing centroid-based routing in maximum inner product search (MIPS) with learned support functions. The work argues that centroid-based routing can miss nearest neighbors when clusters contain outliers far from their centroids, especially when queries (e.g., short user questions) differ geometrically from keys (e.g., long curated documents).

The researchers frame MIPS as a support function—mapping a query to the largest inner product with any key in a set—and propose training small neural networks (SupportNet and KeyNet) to approximate this function offline for a fixed key database. SupportNet learns the support function directly using input convex neural networks, enabling exact routing by evaluating how well a query matches any key within a cluster. KeyNet learns to output the best-matching key vector, with a geometric consistency loss to keep predictions grounded.

Both models are trained on examples drawn from the relevant query distribution: for each sampled query, ground-truth best matches are precomputed via exhaustive search and stored as supervised data. At inference, SupportNet scores clusters by their support functions, routing queries to the most promising clusters before an exhaustive search within those clusters.

The paper, presented at ICML 2026, reports that this approach reduces retrieval latency while maintaining or improving accuracy compared to exhaustive and approximate baselines, particularly in settings where query distributions are stable and well-characterized.

Sources
  1. 01Apple — Machine Learning Research — Faster Retrieval for Better, More Efficient RAG
Also on Research

Stories may contain errors. Dispatch is assembled with AI assistance and curated by human editors; despite the trust-score filter, mistakes happen. We correct publicly — every article links to its revision history. Nothing here is financial, legal, or medical advice. Verify before relying on any claim.

© 2026 Dispatch. No ads. No sponsorships. No paid placement. Reader-supported via Ko-fi.

Built by a person who cares about honest AI news.