--- title: Faster Retrieval for Better, More Efficient RAG | daily.dev description: Apple researchers propose SupportNet and KeyNet, small neural networks that speed up maximum inner product search (MIPS), the retrieval step behind RAG.... canonical: https://daily.dev/posts/faster-retrieval-for-better-more-efficient-rag-twbymxkf3 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Faster Retrieval for Better, More Efficient RAG | daily.dev og:description: Apple researchers propose SupportNet and KeyNet, small neural networks that speed up maximum inner product search (MIPS), the retrieval step behind RAG.... og:url: https://daily.dev/posts/faster-retrieval-for-better-more-efficient-rag-twbymxkf3 og:image: https://api.daily.dev/og/posts/TWbYMxkF3.png og:image:alt: Faster Retrieval for Better, More Efficient RAG og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Faster Retrieval for Better, More Efficient RAG **[Apple Machine Learning Research](https://daily.dev/sources/mlapple)** · 11 min read · 3 upvotes · 0 comments ## Summary Apple researchers propose SupportNet and KeyNet, small neural networks that speed up maximum inner product search (MIPS), the retrieval step behind RAG. Instead of routing queries with simple centroid comparisons, the networks learn the support function or best-matching key for a fixed database of keys. On Natural Questions with 2.6M passages, SupportNet hits 95% routing recall 1.9x faster than centroid routing (1.74ms vs 3.36ms). KeyNet reaches 99% recall up to 8x faster (0.23ms vs 1.8ms) and needs up to 11x fewer scanned keys. The paper was presented at ICML 2026. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much faster is KeyNet than raw query search for maximum inner product search retrieval? KeyNet reaches 99% recall up to 8 times faster than using the raw query, taking 0.23 ms versus 1.8 ms with the largest model size on the Natural Questions benchmark with roughly 2.6 million passages. It also needs up to 11 times fewer scanned keys to hit that recall level, 0.009% versus 0.10% of the database. _Teams tuning RAG retrieval latency can track research like this on daily.dev._ ### What is the difference between SupportNet and KeyNet for speeding up vector search routing? SupportNet learns a cluster's support function to produce a calibrated score useful for re-ranking or thresholding, while KeyNet directly predicts the best-matching key vector and works as a query rewriter in front of an existing IVF index. KeyNet suits cases where inference speed is the bottleneck; SupportNet suits cases needing a calibrated score or where compute is tight at small model sizes. _Developers choosing between retrieval routing strategies can follow comparisons like this on daily.dev._ ### Why does centroid-based routing miss the correct cluster in approximate nearest neighbor search? Centroid routing compares a query only to each cluster's representative centroid, so it cannot account for the spread of keys within a cluster. A cluster can hold a key far from its own centroid but close to the query, and when the query is near such an outlier, centroid routing scores that cluster low and skips it, missing the true best match. _daily.dev helps engineers debugging retrieval recall issues keep up with research on indexing tradeoffs._ ## Similar posts on daily.dev - [Proxy-Pointer RAG: Structure Meets Scale at 100% Accuracy with Smarter Retrieval](https://daily.dev/posts/proxy-pointer-rag-structure-meets-scale-at-100-accuracy-with-smarter-retrieval-mhzxjuu51) · Towards Data Science · 2 upvotes · 0 comments - [Knowledge graph RAG: structured retrieval for AI agents](https://daily.dev/posts/knowledge-graph-rag-structured-retrieval-for-ai-agents-k6cymxiik) · Redis · 1 upvotes · 0 comments - [Why production RAG systems give confident, wrong answers at scale](https://daily.dev/posts/why-production-rag-systems-give-confident-wrong-answers-at-scale-7b6euvauc) · The New Stack · 0 upvotes · 0 comments --- Tags: [#apple](https://daily.dev/tags/apple), [#rag](https://daily.dev/tags/rag), [#neural-networks](https://daily.dev/tags/neural-networks), [#vector-search](https://daily.dev/tags/vector-search) [View this post on daily.dev](https://daily.dev/posts/faster-retrieval-for-better-more-efficient-rag-twbymxkf3) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Faster Retrieval for Better, More Efficient RAG","url":"https://daily.dev/posts/faster-retrieval-for-better-more-efficient-rag-twbymxkf3","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/faster-retrieval-for-better-more-efficient-rag-twbymxkf3"},"datePublished":"2026-11-05T17:37:11.711Z","dateModified":"2026-11-05T18:02:47.899Z","description":"Apple researchers propose SupportNet and KeyNet, small neural networks that speed up maximum inner product search (MIPS), the retrieval step behind RAG....","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8dda20dea5f9a8a0e25a1b7bb36a1a20?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8dda20dea5f9a8a0e25a1b7bb36a1a20?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Apple Machine Learning Research","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Apple Machine Learning Research","logo":"https://media.daily.dev/image/upload/s--mD-1uSe0--/f_auto,q_auto/v1785654191/logos/mlapple","url":"https://daily.dev/sources/mlapple"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/faster-retrieval-for-better-more-efficient-rag-twbymxkf3","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"apple,rag,neural-networks,vector-search","timeRequired":"PT11M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Apple Machine Learning Research","item":"https://daily.dev/sources/mlapple"},{"@type":"ListItem","position":3,"name":"Faster Retrieval for Better, More Efficient RAG"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/faster-retrieval-for-better-more-efficient-rag-twbymxkf3#faq","mainEntity":[{"@type":"Question","name":"How much faster is KeyNet than raw query search for maximum inner product search retrieval?","acceptedAnswer":{"@type":"Answer","text":"KeyNet reaches 99% recall up to 8 times faster than using the raw query, taking 0.23 ms versus 1.8 ms with the largest model size on the Natural Questions benchmark with roughly 2.6 million passages. It also needs up to 11 times fewer scanned keys to hit that recall level, 0.009% versus 0.10% of the database. Teams tuning RAG retrieval latency can track research like this on daily.dev."}},{"@type":"Question","name":"What is the difference between SupportNet and KeyNet for speeding up vector search routing?","acceptedAnswer":{"@type":"Answer","text":"SupportNet learns a cluster's support function to produce a calibrated score useful for re-ranking or thresholding, while KeyNet directly predicts the best-matching key vector and works as a query rewriter in front of an existing IVF index. KeyNet suits cases where inference speed is the bottleneck; SupportNet suits cases needing a calibrated score or where compute is tight at small model sizes. Developers choosing between retrieval routing strategies can follow comparisons like this on daily.dev."}},{"@type":"Question","name":"Why does centroid-based routing miss the correct cluster in approximate nearest neighbor search?","acceptedAnswer":{"@type":"Answer","text":"Centroid routing compares a query only to each cluster's representative centroid, so it cannot account for the spread of keys within a cluster. A cluster can hold a key far from its own centroid but close to the query, and when the query is near such an outlier, centroid routing scores that cluster low and skips it, missing the true best match. daily.dev helps engineers debugging retrieval recall issues keep up with research on indexing tradeoffs."}}]} ```