Impact of Vector Search
summary
Claims that purpose-built vector databases beat bolt-on adapters for scale and latency, with concrete numbers and tools: HNSW graph index and libraries like nmslib and annoy, vector sizes (1536d ≈ 6 KB so 1M vectors ≈ 6 GB (1536d)) and support for very high dims (claims up to 65,000). Shows operational techniques—64× quantization to save RAM, one-call hybrid retrieval and a live 1 GB free cluster experiment—so the talk favors measurable tradeoffs over marketing.
AI-written summary · about
chapters (13)
- 0:00Vector search impact
- 2:32Unstructured data in AI
- 6:04Lexical to semantic search
- 9:57Purpose-built vector DBs
- 12:27What defines a vector database
- 16:42Vector search performance
- 21:35Vector DB innovation
- 24:27Multi-vector retrieval
- 27:43Summary and next steps
- 28:04Q&A partitioning
- 30:53PG Vector for 100 million vectors
- 31:55Quadrant vs others
- 33:37Why Quadrant over Elasticsearch
related talks
39mWhat's in a Vector Database?
34mThe Challenges of Making Vector Search Billion-scale
44mVector Search Beyond the Hype: Real-World Implementation and Performance Insights
01h 23mVector Databases Demystified: Discover how they work
26mThe Rise of Vector Databases - Lessons from the Milvus Community
43mOpenSearch and the Path to a Trillion-scale Vector Database
53mVectors: how to better support a nasty data type in PostgreSQL
32mDistributed Vector Databases - What, Why, and How