Faster retrieval for efficient RAG
The gist
Apple researchers published KeyNet and SupportNet, neural networks that route queries to their best-matching key.
The method cuts time to 99% recall by up to about 8×.
How it unfolded
- Nov 5, 2026Apple publishes "Faster Retrieval for Better, More Efficient RAG," describing KeyNet and SupportNet, which route queries toward their best-matching key; KeyNets of XS, S, M and L sizes equal 1%, 5%, 10% and 20% of the database size.
- Nov 5, 2026Apple reports reaching 99% recall takes up to about 8× less time with the largest model (0.23 ms vs 1.8 ms for the raw query), and scanning just 0.001% of the database retrieves the true key 89–97% of the time.
- Nov 5, 2026Marco Cuturi posts on LinkedIn about the blog post, thanking Peter Gray and his team and crediting Joao Monteiro, Michal Klein and Theo Olausson on the project.
Key terms
- RAG
- A retrieval-augmented generation approach in which an AI agent autonomously retrieves relevant external information to guide its responses.
- maximum inner product search
- A search problem and class of algorithms that maximize the inner product between a query and stored data items.
Source: Wikipedia