Why Database Vector Indexes Need Hierarchical Navigable Small Worlds
from Database Tech with Fexingo: SQL, NoSQL, and Data Storage Conversations Hosted by Fexingo
Vector databases are everywhere now, but most teams slap an HNSW index on their embeddings without understanding why it works or when it breaks. Lucas and Luna drill into the Hierarchical Navigable Small World graph structure — how it layers neighbor graphs from coarse to fine, why the search time scales logarithmically instead of linearly, and the one tuning parameter that kills recall if you set it wrong. They walk through a real scenario: a semantic search pipeline at a mid-size e-commerce company that saw recall drop from 92 percent to 63 percent overnight because of a single efConstruction misconfiguration. No hype, just the internals: how HNSW navigates, what the 'small world' property actually guarantees, and why brute-force re-ranking still matters at the leaf level. If you're using pgvector, Pinecone, or Weaviate and you've ever wondered what's happening inside that index call, this episode is for you. #VectorDatabases #HNSW #HierarchicalNavigableSmallWorld #Embeddings #VectorSearch #ApproximateNearestNeighbor #ANN #pgvector #Pinecone #Weaviate #SimilaritySearch #DatabaseIndexes #SemanticSearch #RecallVsLatency #Technology #FexingoBusiness #BusinessPodcast #DatabaseTech Keep every episode free: buymeacoffee.com/fexingo
Embed this episode
No transcript yet.