ScreenAI: A Vision-Language Model for UI and Infographics Understanding

Abstract

The k-nearest neighbor (kNN) query is a cornerstone of similarity-based applications across various domains. While prior work has enhanced kNN search efficiency, it typically focuses on approximate methods for high-dimensional data or exact methods for low-dimensional data, often assuming static query and data distributions. This creates a significant gap in accelerating exact kNN search for low-to-medium dimensional data with dynamic query distributions. To fill this gap, we propose App2Exa, a cache-guided framework that integrates approximate and exact kNN search. App2Exa utilizes a dynamically maintained cache graph index to retrieve approximate results, which subsequently guide exact search using a VP-Tree with a best-first strategy. A benefit-driven caching mechanism further optimizes performance by prioritizing vectors based on frequency, recency, and computational cost. Experimental results demonstrate that App2Exa significantly boosts efficiency, providing a robust and scalable solution for evolving query patterns and enabling exact kNN search to support higher dimensionality more effectively.