Vector search
Search powered by vector similarity — finding the K nearest embedding vectors to a query vector, typically using approximate-nearest-neighbor algorithms like HNSW or IVF.
Vector search is what makes semantic retrieval practical. Documents and queries get encoded as embedding vectors; "relevance" becomes "how close is your query vector to a document vector in high-dimensional space." Approximate-nearest-neighbor algorithms make this fast even over 100M+ vectors.
The dominant index types in 2026: HNSW (graph-based, accurate, memory-hungry), IVF (clustered, scalable, slightly less accurate), and binary indexes for billion-scale workloads. Most production stacks use HNSW for under 100M vectors and switch to compressed indexes past that.
For agents, vector search rarely runs alone. The 2026 pattern is hybrid retrieval: combine vector search with keyword/BM25 search, fuse the scores, and rerank the top results with a stronger model. This consistently beats pure-vector retrieval on production RAG benchmarks.
Frequently asked
What is the difference between vector search and semantic search?+
Semantic search is the user-facing capability ("find content that means the same thing"). Vector search is one implementation technique. Most modern semantic search systems are vector-search-powered, but you can do semantic search with cross-encoders or LLM judges too.
How accurate is vector search?+
Recall@10 of 90–98% is typical for HNSW on production indexes. The accuracy depends more on the embedding model and chunking strategy than on the index algorithm.