Hierarchical navigable small world

En étudiant l'article Wikipedia "Base de données vectorielle", je découvre la liste de différents algorithmes Approximate Nearest Neighbor.

#JaiDécouvert feature extraction algorithms.

These feature vectors may be computed from the raw data using machine learning methods such as feature extraction algorithms, word embeddings or deep learning networks. The goal is that semantically similar data items receive feature vectors close to each other.

source

J'apprends :

In recent benchmarks, HNSW-based implementations have been among the best performers.

source

Je lis :

Databases that use HNSW as search index include:

• Apache Lucene Vector Search
• Chroma
• Qdrant
• pgvector
• Clickhouse
• …

source

En interrogeant Claude Sonnet 4, j'apprends :

Benchmark indicatif (1M vecteurs 768D) :

Métrique	Qdrant	pgvector	Elasticsearch
Temps indexation	15 min	45 min	25 min
Requête/sec	2000+	500-800	800-1200
RAM utilisée	4 GB	6 GB	8 GB+
Précision @10	0.95	0.92	0.94
Date création	2021	2021	2022 (support HNSW)
Langage	Rust	C	Java
	Open Source	Open Source	Open Source

Claude Sonnet 4