Ilmu Komputer & AI editorial
Open AccessOA2026
CLIP: Lightweight Cosine-Law-Based Inverted-List Pruning for IVF-Based Vector Search
A cosine-law pruning technique enabling inter- and intra-cluster pruning with
cluster elimination and logarithmic vector filtering, achieving up to 78% pruning and 141% throughput improvement.
Yitong Song; Shuhang Lu; Pengcheng Zhang; Jianliang Xu· 2026· DOI 10.48550/arXiv.2606.29968
The core problem
Vector search is a core component of modern multimodal retrieval systems. Inverted file (IVF)-based methods are widely adopted due to their scalability, efficient updates, and hardware friendliness. However, they suffer from coarse-grained execution: each query probes many clusters and exhaustively scans all vectors within them, leading to high query latency. Prior pruning strategies mitigate this but incur substantial overhead, lack cluster-level pruning, and compromise update efficiency due to heavy metadata maintenance. This paper proposes CLIP, a lightweight cosine-law-based pruning technique that supports both inter- and intra-cluster pruning, substantially reducing unnecessary cluster and vector accesses with negligible overhead.
Innovation
Extensive experiments show that CLIP variants achieve up to 78% pruning and 69% higher efficiency over static IVF baselines. Specifically, IVF-CLIP and HIVF-CLIP reduce the number of distance computations and cluster accesses significantly, leading to lower query latency. For dynamic workloads, LSM-IVF improves throughput by up to 141% over dynamic IVF baselines with comparable update efficiency. The pruning overhead is negligible, as the lower bound computations are lightweight and can be performed quickly. The analytical guarantee ensures that no true nearest neighbors are missed, maintaining high recall.
Vector search is a core component of modern multimodal retrieval systems. Inverted file (IVF)-based methods are widely adopted due to their scalability, efficient updates, and hardware friendliness. However, they suffer from coarse-grained execution: each query probes many clusters and exhaustively scans all vectors within them, leading to high query latency. Prior pruning strategies mitigate this but incur substantial overhead, lack cluster-level pruning, and compromise update efficiency due to heavy metadata maintenance. This paper proposes CLIP, a lightweight cosine-law-based pruning technique that supports both inter- and intra-cluster pruning, substantially reducing unnecessary cluster and vector accesses with negligible overhead.
CLIP exploits the monotonicity of cosine-law-based lower bounds to eliminate undesirable clusters in
time and filter batches of irrelevant vectors in logarithmic time in the list size, with a tight analytical guarantee. The core idea is to compute a lower bound on the cosine similarity between a query and any vector in a cluster, based on the cluster centroid and the maximum distance from the centroid to its vectors. If this lower bound is less than the current top-k threshold, the entire cluster can be pruned. Formally, for a query , cluster centroid , and cluster radius (maximum distance from to any vector in the cluster), the lower bound on the cosine similarity is:
Why it matters
The CLIP technique addresses key limitations of existing IVF-based methods by providing a lightweight, effective pruning mechanism that operates at both cluster and vector levels. The
cluster elimination and logarithmic vector filtering enable significant speedups without compromising accuracy. The hierarchical extension in HIVF-CLIP allows adaptive probing, further improving efficiency for skewed data distributions. The LSM-IVF design demonstrates that CLIP can be integrated into dynamic indexing structures, offering fast updates and high query throughput. The results indicate that CLIP is a promising approach for large-scale vector search systems, with potential for further optimizations in metadata management and parallelization. Future work could explore combining CLIP with other indexing techniques and extending it to other distance metrics.
Who should read this
CS practitioners and researchers
Opening member content…