Apache Lucene Vector Search Update
summary
Survey of recent Lucene vector work that lists concrete wins and how to get them: scalar quantization now supports 1-bit documents / 4-bit queries and an optimized scalar codec in Lucene 10.4 for much lower storage with acceptable recall; segment-concurrent search plus reentrant prorating safely reduces per-segment K work; filtered search uses acorn filtering to walk only matching graph neighborhoods; merges reuse existing HNSW graphs via covering seeds to cut reindexing CPU, and small flushes skip graph building to save indexing cost. JDK Vector API and off-heap memory improvements enable the performance gains.
AI-written summary · about
chapters (13)
- 0:00Apache Lucene vector search
- 1:42What is Lucene
- 3:25Quantization in vector search
- 10:20Radius search and new queries
- 12:26Concurrent search in Lucene
- 14:35Diversifying KNN and filtering
- 17:25Seeded vector search and segments
- 22:45Indexing optimizations
- 29:45Merging policies and graph reuse
- 31:32Java performance for vector search
- 37:05Performance testing and JDK upgrades
- 38:50Closing remarks
- 39:32Q&A
related talks
40mIntroducing Multi-valued Vector Fields in Apache Lucene
52mBeyond Single Vectors: Multi-Vector Search for Enhanced...
41mKeynote: From Keyword Search to Vector Powered Insights: Th...
39mWhat's in a Vector Database?
47mBoosting AI With OpenSearch: Vector Stores, Similarity Search, and Retrieval-Augme...
41mPractical Vector Search Configuration, Tuning, RAG, and the Road Ahead
44mVector Search Isn’t Enough
33mEnter the Brave New World of GenAI with Vector Search