Abstract
<title>Abstract</title> <p>Efficient similarity search in high-dimensional vector spaces has become a core component of many modern AI applications, such as semantic search, recommendation systems, fraud detection, and embedding-based large language models. As data volumes grow, selecting an appropriate indexing strategy becomes critical for balancing accuracy, query latency, and memory usage. Facebook AI Similarity Search offers a diverse set of index structures designed to address these trade-offs under different workload conditions. This study provides a structured experimental evaluation of several widely used Facebook AI Similarity Search index types: IndexFlatL2, IndexFlatIP, IndexIVFFlat, IndexIVFPQ, IndexIVFSQ, and IndexHNSW. The analysis investigates their computational behavior, recall performance, and sensitivity to various distance metrics, including L1, L2, inner product, and cosine similarity. It also explores the effects of quantization techniques, clustering strategies, and graph-based connectivity parameters on overall efficiency. The results highlight clear differences in performance patterns across index families and offer practical guidance for selecting appropriate configurations in real-world, large-scale deployment scenarios.</p>