The Challenges of Making Vector Search Billion-scale
summary
Claims concrete tradeoffs and engineering wins: Qdrant uses HNSW with segmenting and memory-mapped files to avoid full in‑RAM graphs, demonstrates 100 megabytes for a million vectors in one case, and applies scalar and binary quantization to cut storage (about 75% per scalar quant and up to 32x memory reduction for some high‑dim embeddings) while rescoring on original floats to recover quality; benchmarks show ~98.8% precision and real deployments at billion‑vector scale.
AI-written summary · about
chapters (10)
- 0:00Vector search and scaling
- 2:12Semantic vs keyword search
- 7:14Scaling with approximate neighbors
- 10:55Qdrant for billion-scale search
- 17:55Limitations of vector search
- 21:07Memory vs performance trade-offs
- 25:05Optimizing memory usage
- 28:59Distributed Qdrant search
- 30:07Quantization for memory efficiency
- 32:57Q&A
related talks
35mImpact of Vector Search
39mWhat's in a Vector Database?
44mVector Search Beyond the Hype: Real-World Implementation and Performance Insights
43mOpenSearch and the Path to a Trillion-scale Vector Database
40mUnder the hood of vector search with JVector
36mWhen Semantic Search Breaks in your RAG systems: RAM Walls, Silent Failures, and the Architecture Decisions
32mUnder the Hood: Key Algorithms Powering Vector Search
54mDissimilarity search...