Fast and Explainable Search on a Budget With OpenSearch Neural Sparse
summary
Claims that doc-only inference-free neural sparse retrieval (token-weight vectors) yields high relevance with predictable CPU latency: OpenSearch models on Hugging Face hit 1M+ downloads/month and ship as index-time sparse vectors so queries skip expensive inference. Benchmarks on ~9M docs show index sizes ~1GB (BM25), ~7GB (neural sparse) vs ~60–65GB (dense), and latencies near BM25 (~40ms) with two‑phase (~45ms) or ~11ms using Sysmaker at billion scale. Shows end-to-end setup with OpenSearch semantic field, Sentence Transformers encoders, mappings, explainable token-weight evidence and copy‑paste configs.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:42Lexical search limits
- 1:24Vector search costs
- 2:08Learned sparse retrieval
- 3:54Inference-free vs bi-encoder
- 4:58Sparse vs dense efficiency
- 6:24Neural sparse overview
- 7:50Billion-doc benchmarks
- 8:53Using sparse models
- 10:41Product search demo
- 11:47BM25 comparison
- 12:29Q&A
related talks
32mBudget Friendly Semantic Search With Neural Sparse Search
41mFine-tuning Neural Sparse Model for Domain Specific Data From an Existing OpenSearch Index
46mNatural Language & Search, Improve your Search w/ Vector capabilities
48mUnlocking Next-gen Semantic Search With OpenSearch 3.x
30mVector Search Made Simple: Getting Started with OpenSearch for AI Applications
32mRetrieval Augmentation and Semantic Search at Scale
47mWhy, What, When and How of Running Diff Types of Sear...
33mBeyond Benchmarks: Real-World AI Search Performance in OpenSearch (Lucene Vs Faiss Vs JVector)