HNSW vs. LSH: How Elasticsearch hits 0.99 recall@10 at 15,000 QPS — and what it costs

This article explains how Elasticsearch achieves high-performance vector search using HNSW indexing, reaching 15,000 QPS with a 0.99 recall@10 for float32 vectors. It highlights how implementing DiskBBQ quantization can further boost throughput to 55,000 QPS with only a minimal reduction in recall by significantly reducing memory overhead. Ultimately, the text explores the technical trade-offs between search accuracy, speed, and memory usage when managing high-dimensional datasets.