SW-graph Won the First ANN-benchmarks in 2015 and Was Already Production-Ready, Author Recalls
srchvrs · x · 2026-09-11
srchvrs recounts a 2015 anecdote in vector search: after Google's Sujith Ravi presented a large-scale distributed LSH system, he recommended NMSLIB, which already included SW-graph, a pre-HNSW graph-based retrieval algorithm. SW-graph then won the first ANN-benchmarks; though not the inventor, he sped up the original Java port by an order of magnitude. He adds that despite scalability concerns, the 2015 version ran reasonably in production (indexing cost is often small vs. retrieval cost), was already faster than typical LSH, and only trailed FALCON-LSH due to a bug fixed circa 2016.
More from Research
- Distilling Mast3R to work on the Moon: ECCV 2026 submission showcased — ducha_aiki · 2026-09-11
- PathSegmentor Segments Pathology Images from a Single Natural-Language Prompt — bravo_abad · 2026-09-11
- WorldAgents (ECCV 2026): 3D scenes in 25 min for a few euros per model — ducha_aiki · 2026-09-11
- Why Importance Sampling Is Everywhere in LLM RL: The Clipping Logic of PPO and TIS — cwolferesearch · 2026-09-11
- AgileRL Arena v1.0: manifest-driven RL training with LoRA/GRPO finetuning — Balance- · 2026-09-11
- Mike Frank separates truth from hype on "fruit fly learns Mario" claims — MikePFrank · 2026-09-11