Faster Retrieval for Better, More Efficient RAG

Apple researchers, in a paper presented at ICML 2026, propose training small neural networks offline on a fixed key database and a sampled query distribution to route maximum inner product search queries more accurately than centroid-based clustering.