Implementation of filtered-Nearest Neighbour Search Methods for High-Dimensional Vectors

Abstract

Approximate nearest neighbor search (ANNS) has become a fundamental component of modern retrieval-augmented generation (RAG) and recommendation systems. However, combining vector similarity search with filtering on structured attributes poses unique challenges to the standard ANNS formulation. Most available methods today struggle to achieve both high recall and high query throughput under tight filtering constraints. To address the filtered k-ANNS problem, several specialized algorithms have recently been proposed. In this thesis, we implement two state-of-the-art methods—ACORN (ANN Constraint-Optimized Retrieval Network) and IVF2—and integrate them directly into FAISS, a widely used C++ library for high-performance vector search. We evaluate the new implementations across multiple benchmark datasets against industry-standard baselines. Our benchmarks evaluate critical performance dimensions, including index construction time, memory footprint, recall, and query throughput (QPS). Furthermore, we analyze the theoretical guarantees, strengths, and limitations of each method under varying query selectivities. Experimental evaluation demonstrates that ACORN achieves almost the same query throughput for equal recall across most datasets and selectivities as the IVF baseline. Meanwhile, IVF2 achieves a 5× to 20× speedup in QPS compared to standard baselines on datasets with semi-structured attributes. Our open-source integration provides a scalable solution for high-throughput, attribute-constrained vector search in production database systems.