What's in a Vector Database?
summary
Walks through building a usable vector database from an object store to indexes to sharding and replication, demonstrating that a working prototype is possible in ~25 minutes and scales with the right choices. Uses HNSW graphs (tunable M and EF) for fast approximate search, shows raw-memory math (1M vectors ≈ 6 GB, 100M ≈ 650 GB), uses product quantization and binary quantization to cut size ~60–70% with over‑fetch+rescore, adds roaring bitmaps for fast filtering, and combines vector and keyword scores via fusion (RRF or normalized sum).
AI-written summary · about
chapters (20)
- 0:00Vector databases intro
- 1:06Why trending now
- 2:33What vectors represent
- 4:20Building from scratch
- 5:46Scaling search with indexes
- 7:53Flat vs graph indexes
- 9:38HNSW indexes and parameters
- 12:05Memory challenges and quantization
- 15:14Combining indexes for hybrid search
- 18:44Scaling with sharding
- 21:12Hardware planning
- 22:37Developer experience
- 25:50Key architecture summary
- 26:33Resources
- 28:25Q&A: Combining vector and keyword search
- 30:23Q&A: Setting similarity thresholds
- 32:10Q&A: Hybrid search trade-offs
- 33:42Q&A: Quantization impact on recall
- 35:48Q&A: Handling multimodal data
- 37:13Q&A: Choosing the right DB
related talks
35mImpact of Vector Search
01h 23mVector Databases Demystified: Discover how they work
34mThe Challenges of Making Vector Search Billion-scale
41mGetting Started With Vector Databases
26mThe Rise of Vector Databases - Lessons from the Milvus Community
48mA Fun & Absurd Introduction to Vector Databases
50mA fun and absurd introduction to Vector Databases
46mA fun and absurd introduction to Vector Databases