Tutorial on Approximate Nearest Neighbor Search (ANNS) – Techniques and Open Problems

1 ANNS AND RECENT DEVELOPMENTS Approximate nearest neighbor search (ANNS) is a classical algorithmic problem that is increasingly relevant in practice today across a variety of AI application domains. For AI-first applications, ANNS is the key index that connects neural networks for search [27], recommendation [6] and content generation/summarization [10, 18, 19, 25, 26] with relevant items in the knowledge store. This linking process helps better ground the output of the models and generates higher quality results. While ANNS is a well-studied problem, classical algorithms like LSH [2] and its derivatives fail to meet the requirements of this new generation of applications. Some of the shortcomings include inadequate query performance for large indices of high dimensional vectors (107 to 1012 vectors with dimensions in the range of 102 − 104) and memory-inefficiency and difficulty scaling in the external-memory setting. Perhaps most importantly, existing ANNS approaches are difficult to adapt to new and natural feature requirements motivated by emerging applications. Examples include (a) designing streaming indices, (b) designing ANN indices that can support a combination of hard matches and nearest neighbor search [34], and (c) designing indices that can quickly answer filtered queries. Addressing these new requirements has been a productive and active new research area, with numerous research papers [4, 15, 20, 22– 24, 31, 32] and open search packages [11, 12, 22, 28] aiming to push the research and development frontier. There are also vector-searchas-a-service companies (e.g. Milvus, Pinecone, Qdrant, Vespa, Weaviate, Zilliz) that package this research for commercial use, not to mentioned extensively customized in-house solutions at larger companies where several products stand on ANNS.