Under the hood of vector search with JVector
summary
JVector is a Java embedded library that implements graph-based ANN with PQ rerank to run indexes in memory and on SSD, aiming for high recall and low latency. It adopts single-level Vamana-style graphs, product quantization + ADC for queries, and reranks full-resolution vectors to recover quality — e.g. 64x compression with 3x overquery recovers relevance. Binary quantization and “finger” tricks are explored but underperform; inserts are lock-free and scale with cores, and on-disk layout targets predictable random reads for larger-than-memory sets.
AI-written summary · about
chapters (5)
- 0:00Introduction and motivation
- 13:24Vector search fundamentals
- 27:13JVector design and goals
- 37:06Takeaways and future work
- 38:34Q&A
related talks
44mVector Search Beyond the Hype: Real-World Implementation and Performance Insights
34mThe Challenges of Making Vector Search Billion-scale
32mUnder the Hood: Key Algorithms Powering Vector Search
01h 09mVector Search: Ask Me Anything!
33mEnter the Brave New World of GenAI with Vector Search
54mDissimilarity search...
35mImpact of Vector Search
11mCassandra Vector Search Performance Tuning