Distributed Vector Databases - What, Why, and How
Talkat Data on Kubernetes Day North America 2023 · on Monday, 2023-11-06· at 10:25· W471ab, McCormick Place·
summary
Explains how unstructured data turns into 512-dimensional vectors via pretrained embedding models (OpenAI, Hugging Face), how those vectors live in memory-heavy indices and are searched with an HNSW index for approximate nearest neighbors, and how retrieval augmented generation (RAG) stitches nearest-documents into LLM prompts. Advises running distributed vector stores on Kubernetes with local NVMe and lots of RAM; notes practical numbers (example: 1M embeddings ≈6GB RAM), tradeoffs of vector length versus accuracy, and tooling like PGVector, LlamaIndex and LangChain.
AI-written summary · about
chapters (16)
- 0:00Introduction to vectors
- 1:14Understanding vector data
- 3:42Tokens and embedding models
- 5:33Choosing an embedding model
- 7:22Vector space and clustering
- 9:25Querying the database
- 11:12HNSW indexing algorithm
- 14:31HNSW layers and applications
- 17:58Use cases and benefits
- 19:27Clear boundaries for teams
- 21:33Practical use cases
- 22:28Retrieval augmented generation
- 24:49Vector database architectures
- 25:42Infrastructure requirements
- 27:33Summary and conclusion
- 28:29Q&A session
related talks
01h 23mVector Databases Demystified: Discover how they work
39mWhat's in a Vector Database?
06mWhat makes vector databases essential for RAG and AI apps?
45mA gentle introduction to Vector Databases
33mVector Databases for GenAI applications
01h 46mVector Databases (for GenAI & beyond)
50mA fun and absurd introduction to Vector Databases
35mImpact of Vector Search