Faster Retrieval for Better, More Efficient RAG
Web page · https://machinelearning.apple.com/research/faster-retrieval
First seen by Oriel on .
AI summary
Apple researchers present amortized MIPS, training small neural networks on known query distributions to accelerate and improve retrieval over fixed key databases for RAG and other systems.
- SupportNet approximates per-cluster support functions to route queries by the best key in each cluster, rather than by centroid proximity.
- KeyNet predicts the best-matching key vector and rewrites queries before an existing IVF index, improving search without replacing the index.
- On Natural Questions with 2.6M passages and 128 clusters, SupportNet reached 95% routing recall 1.9× faster than centroid routing.
- KeyNet reached 99% recall in 0.23 ms versus 1.8 ms for raw queries, while scanning 0.001% of the database achieved up to 96.7% retrieval.
- KeyNet favors inference speed, while SupportNet provides calibrated scores and remains competitive for small models; both require offline training for fixed keys and query distributions.
#mips #rag #neural retrieval #ivf
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hacker News: 237, latest observation ; +222 over observed period since
Comments
Comment count reported by the source.
- Hacker News: 39, latest observation ; +39 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hacker News | 50455633 | https://machinelearning.apple.com/research/faster-retrieval Read discussion |