HNSW

HNSW, short for Hierarchical Navigable Small World, is a graph-based ANN algorithm introduced in a 2016 paper by Malkov and Yashunin that has become the dominant index type in modern vector databases. The algorithm builds a layered graph where each layer is a subset of the layer below, with the top layer containing a few highly-connected nodes that serve as long-distance shortcuts. Query traversal starts from the top layer and descends greedily, refining the candidate set at each layer until it converges on the nearest neighbors. HNSW achieves excellent recall-vs-latency trade-offs, typically delivering Recall@10 above 95% with millisecond latencies on million-scale vector collections. Key tuning parameters include M (graph connectivity), efConstruction (build-time search effort), and ef (query-time search effort), each balancing accuracy against memory and time. Pinecone, Weaviate, Qdrant, Milvus, pgvector, Elasticsearch, OpenSearch, and Redis all default to HNSW or offer it as a primary option, making it the most operationally proven ANN algorithm in production AI governance environments.

HNSW tuning with Centralpoint: Centralpoint stays model-agnostic across whatever HNSW implementation you use, metering retrieval-plus-generation tokens so finance sees the actual cost-quality trade-off. Prompts and skills stay on-premise, and chatbots backed by HNSW retrieval embed across portals with one line of JavaScript and full audit logs.


Related Keywords:
HNSW,,