Reproduce “Aker: Density-Aware Approximate Caching for Vector Search (Extended Version)” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sukjoon Oh, Minki Kang, Dohyun Kim, Baotong Lu, Jing Liu, Qianxi Zhang, Qi Chen, Youjip Won
Aker, developed by KAIST and Microsoft Research, introduces a density-aware approximate caching system for vector search that dynamically adjusts per-query similarity thresholds and employs a "del-consistency" model for robust updates. This system maintained high recall while significantly boosting query throughput by up to 3.2x and reducing read amplification, even under continuous data modifications.
Key insights
- High-dimensional vector spaces exhibit non-uniform densities, making global similarity thresholds ineffective for balancing accuracy and efficiency in approximate caching; per-query adaptive thresholds are essential.
- A practical cache consistency model for ANNS can be achieved by prioritizing validity (immediate deletion reflection) while tolerating relaxed visibility (lazy insertion handling), significantly reducing refresh overhead.
- Efficient memory management through neighbor object sharing and a dual-index eviction strategy for the query filter are critical for scalable approximate caching in dynamic, disk-based vector search systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.