Vector Search Beyond the Hype: Real-World Implementation and Performance Insights
summary
Shows concrete outcomes from building the jVector KNN plugin for OpenSearch: inline vectors reduce random IO and often lower latency versus aggressive quantization, product quantization enables building billion‑vector graphs on small machines but can cut recall so reranking often cancels IO gains, demonstrates a hybrid HNSW layer that fixed a customer’s pathological distributions, introduces snapshot‑based coordination for concurrent graph inserts and emphasizes tuning threads to CPU SIMD/CMD cores for real speed.
AI-written summary · about
chapters (14)
- 0:00Vector search realities
- 1:45Why we built JVector
- 4:38Dense vectors challenges
- 6:24Recall versus speed
- 10:39Graph index limitations
- 16:01IO bottleneck in search
- 19:14Disk-based nearest neighbor trade-offs
- 21:40Inline vectors performance
- 23:43Quantization impact
- 30:54Graph index challenges
- 33:43Quantization for efficiency
- 37:00Accelerating index construction
- 41:12Testing concurrent construction
- 43:19Search architecture optimizations
related talks
40mUnder the hood of vector search with JVector
35mImpact of Vector Search
34mThe Challenges of Making Vector Search Billion-scale
01h 09mVector Search: Ask Me Anything!
33mBeyond Benchmarks: Real-World AI Search Performance in OpenSearch (Lucene Vs Faiss Vs JVector)
32mUnder the Hood: Key Algorithms Powering Vector Search
41mApache Solr 10: What's Coming up for Vector Search
01h 23mVector Databases Demystified: Discover how they work