OpenSearch and the Path to a Trillion-scale Vector Database
summary
OpenSearch adds concrete engineering: disk-optimized vector search uses binary quantization + HNSW to scan compressed bits, then rescoring on full-precision vectors to keep >90% recall at P90 ≈200 ms and cut cost to ~1/3; quantization options include FP16, int8 and PQ for up to ~64× compression and ~75% scalar savings; GPU index acceleration offloads KAGRA builds to GPUs and converts to HNSW for 10× faster index builds, SIMD (AVX‑512/NEON) boosts search throughput ~3×, and hybrid/reciprocal-rank fusion improves NDCG@10 by 8–12%.
AI-written summary · about
chapters (7)
- 0:00OpenSearch vector database intro
- 7:21OpenSearch vector DB evolution
- 13:24Memory challenges at billion scale
- 24:46Hybrid lexical and vector search
- 29:21Path to trillion-scale vectors
- 38:11Future trillion-scale innovations
- 42:47Q&A
related talks
26mOpenSearch and the Path to a Trillion-s...
48mUnlocking Next-gen Semantic Search With OpenSearch 3.x
55mVector Search Made Simple: Getting Started with OpenSearch for AI Applications
30mVector Search Made Simple: Getting Started with OpenSearch for AI Applications
22mVector Search Made Simple: Getting Started With OpenSearch for AI Applications
34mThe Challenges of Making Vector Search Billion-scale
32mRetrieval Augmentation and Semantic Search at Scale
35mImpact of Vector Search