Vector Ops: run vector embedding-powered apps in prod
summary
Argues vectors capture nuance lost to words and shows a working pattern: concatenate content and collaborative features into single vectors, store user and item vectors in a vector database, and run approximate nearest neighbor search for retrieval instead of fetching 10,000 candidates and reranking. Reports thousands QPS per machine, 800–2,000 dims typical, and target recall ~80–90%. Describes an MVP path using Hugging Face or OpenAI embeddings, brute-force or ANN libs for prototyping, realtime vector updates, evaluation metrics and a query manager layer for tweaks, diversity and post-filtering.
AI-written summary · about
chapters (23)
- 0:00intro
- 0:22preface
- 0:36building vector-powered apps
- 1:35what did we lose with language?
- 4:14natural language is a bottleneck
- 4:50natural language is ambiguous
- 5:56vectors are better! (mostly)
- 7:55search before vectors
- 10:32recommendations before vectors
- 14:42search & recommendations with vectors
- 17:40approximate nearest neighbours?
- 20:23it can be measured
- 22:57building the content vectors...
- 25:39and user vectors in the same space!
- 27:53let ann do the heavy lifting
- 29:39query manager on top
- 32:36what will you need to get started & MVP?
- 38:26towards a #vectorops platform
- 42:57what about generative ai?!
- 43:41chat and chains!
- 45:30what the frig?
- 46:42agents & memory!
- 48:41let's connect!
related talks
50mA fun and absurd introduction to Vector Databases
48mA Fun & Absurd Introduction to Vector Databases
39mWhat's in a Vector Database?
01h 46mVector Databases (for GenAI & beyond)
36mUnleashing the Power of Vector Databases for AI-Driven Applications
01h 23mVector Databases Demystified: Discover how they work
48mA fun and absurd introduction to Vector Databases
57mEnter The Brave New World of GenAI with Vector Search