Faster Retrieval for Better, More Efficient RAG
Reported by Apple Machine Learning Research1 source ↓
Apple researchers describe neural networks trained offline to route maximum inner product search queries toward matching database keys. The approach is designed to speed retrieval using patterns in the expected query distribution.
Key points
- The method trains small neural networks for a fixed database of keys
- Training uses examples from the expected query distribution and exhaustive-search matches
- SupportNet estimates support functions; KeyNet predicts the best-matching key
Details
The work, “Amortizing Maximum Inner Product Search with Learned Support Functions,” was presented at ICML 2026, Apple’s machine learning research team reports. It targets maximum inner product search, used in retrieval-augmented generation, recommendation systems and search, where exhaustive comparison becomes costly as databases and vector dimensions grow.
The networks are trained offline for a particular key database, using sampled queries paired with their best matches found through exhaustive search. This lets the models learn patterns in the query distribution rather than relying only on the geometry of the keys; the researchers say the approach trades this upfront training cost for faster, more accurate retrieval at inference time.
The team describes two model types. SupportNet estimates the support function and can be used to improve routing among data clusters, while KeyNet directly predicts the best-matching key vector.
Written from coverage as of
Images
- Apple Machine Learning Research
- Apple Machine Learning Research
- Apple Machine Learning Research
- Apple Machine Learning Research
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh