Scaling Semantic Search to Billions with Neural Sparse ANN
summary
OpenSearch 3.3 adds a neural sparse ANN implementation using the SEISMIC algorithm that clusters posting lists and prunes blocks at query time. It indexes with a sparse_vector mapping (method=sysmic) and benchmarks on 1.29 billion C4 docs with MS MARCO queries show recall@10 90% / 11 ms average latency. Query hyperparameters are top_n, heap_factor, and approximate_threshold; force-merge and segment sizing drive when clustering runs. Cost example: SEISMIC needs ~15 nodes ($35/hr) versus HNSW’s ~42 nodes ($75/hr) for similar recall.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:03Real-world retrieval scale
- 2:50Sparse retrieval for agents
- 3:54Neural sparse search
- 5:48Sysmic billion-scale breakthrough
- 6:08Sysmic deep dive
- 8:03Sysmic in OpenSearch 3.3
- 9:31Billion-scale benchmarks
- 10:58Cost vs dense vectors
- 11:46Sysmic hyperparameter tuning
- 13:11Index building tips
- 15:19Q&A
related talks
41mScaling Search and Powering Vector-Based Retrieval With OpenSea...
30mIncrease text search speed & accuracy with Amazon OpenSearch Service
43mEnhancing Query Performance Through Intelligent Data Co-location
31mMaximize Resource Efficiency with Separated Index and Search Workloads
26mSimilarity search: Take your search engine to the next level
19mSemantic Search for the Rest of Us with DuckDB
28mSearch and Sushi; Freshness Counts
29mSolr Sparse Faceting #bbuzz