Guest Lecture - Vikram Srinivasan

​"RAG is not dead - You are doing retrieval wrong"

As context windows stretch past two million tokens and "context graph" architectures promise emergent intelligence, a growing chorus declares RAG obsolete. Meanwhile, a parallel hype cycle insists vector databases are all you need. Both camps are wrong.
RAG remains the most practical architecture for enterprise knowledge work - but only if you abandon the naive version: chunk, embed, nearest-neighbor lookup, pray. Drawing on DeepMind's proof that fixed-dimension embeddings have fundamental retrieval limits, and on lessons deploying AI search across hundreds of terabytes of messy enterprise data, this talk makes three claims: semantic search matters more than ever, but "semantic" doesn't mean "vector-only." Vector databases are a means, not an end. And long context doesn't replace retrieval - the hard part was never fitting more text into the prompt, it was finding the right text.