HNSW = Hierarchical NSW graph
Vectors linked into a layered graph. Search descends from top (few nodes) to bottom (all nodes). O(log N) lookup.
Advertisement
HNSW = Hierarchical NSW graph
Vectors linked into a layered graph. Search descends from top (few nodes) to bottom (all nodes). O(log N) lookup.
Advertisement
Sharding for scale
Beyond ~10M vectors per shard, split. Query fans out to all shards. Merge top-K globally.
Metadata filtering
'Vectors matching my query AND category = X AND date > Y'. Filter first (or post-filter) to avoid scanning irrelevant vectors.
Rerank for accuracy
Approximate NN gives top 100. Cross-encoder reranks top 10 with more precision. Balance recall vs cost.
Storage choices
pgvector for Postgres integration. Pinecone for managed. Weaviate for hybrid. Qdrant for cost. Choose by scale + ops maturity.