Google open-sources GINGER: 1.8x faster vector search with 2.6x less memory

_reachsumit · x · 2026-11-06

Google released GINGER, an open-source C++ library for graph-based approximate nearest neighbor search (arXiv 2611.08306):

Builds on PiPNN proximity-graph construction and TurboQuant; introduces Nested TurboQuant (NTQ), a two-level quantizer enabling low-precision graph navigation plus high-precision reranking in one representation — near floating-point precision with 2.6x less memory.
Compresses proximity graphs up to 2.5x without hurting throughput; combines graph reordering with AdaptiveGap, a new lossless codec based on StreamVByte.
At 95% recall, query throughput averages 1.8x over each dataset's fastest competitor and at least 2.3x over any single library, growing to 2.9x over compact libraries on high-dimensional embeddings; indexing time also leads.

A substantial new option for teams building vector search / RAG infrastructure.

More from Infra