Technology method · wiki:deep
ANN indexes
Approximate nearest-neighbor structures (HNSW, IVF, …) whose cost scales with dimension — Matryoshka truncation shrinks d before the index pays for it.
ANN indexes answer “nearest vectors” without scanning all rows. Distance and graph/memory cost grow with d. When OpenAI emits 1024 instead of 1536/3072, or mxbai serves 512 instead of 1024, the same HNSW parameters usually get cheaper queries — that is the link to Matryoshka truncation.
Catalog packs that own this substrate live under Vector DBs: FAISS, Qdrant, Milvus, pgvector, Weaviate, and peers. This method page names the index concern so Matryoshka essays do not pretend the store invented nested training.
See also
In this taxonomy
_catalog/taxonomy/tech-ann-indexes.md