Towards a deeper understanding of retrieval and vector databases
summary
Practical takeaways: a deduplication pipeline runs on 15 million images using DinoV2 image embeddings into an HNSW-backed vector DB, exposing that prefiltering by metadata collapses performance once >80–90% of nodes are dropped unless metadata indexing reconnects similar metadata clusters; HNSW buys precision and speed at higher RAM cost. A RAG for a medical company uses Ada-2 text embeddings and flat indexing for a few thousand docs, and source-level filtering raises answer accuracy from 78% to 89%. Quantization and partitioning reduce cost without large quality loss.
AI-written summary · about
chapters (13)
- 0:00Vector databases introduction
- 3:58How vector search works
- 8:15Dedicated vs general-purpose
- 10:00Detecting duplicate real estate images
- 14:36Metadata filtering challenges
- 17:50Image duplicate detection models
- 21:06Building RAG for hotels
- 22:31Vector database optimizations
- 23:57Bias in embedding models
- 26:50Q&A: Metadata filtering
- 28:14Q&A: Complex metadata
- 29:18Q&A: Vector search vs embeddings
- 30:44Q&A: Choosing vector database
related talks
32mDistributed Vector Databases - What, Why, and How
39mWhat's in a Vector Database?
50mA fun and absurd introduction to Vector Databases
34mThe Challenges of Making Vector Search Billion-scale
29mDataEngBytes 2023 - SYD-T2-01 - Rene Essomba
01h 23mVector Databases Demystified: Discover how they work
27mUnified AI & Relational Power: Scalable Distributed Vector Search with
36mWhen Semantic Search Breaks in your RAG systems: RAM Walls, Silent Failures, and the Architecture Decisions