Nick Saraev wrote:
“I built RAG systems for a bunch of clients over the years. Then I realized ppl were doing this with plain text search. We switched. Inference time went down by ~9/10. The cost of maintaining the vector DBs and embeddings went down by ~9/10. The algorithm is BM25 and it is from the 1990s.
Boris Cherny, who created Claude Code, said something similar: plain search outperformed every other approach by a lot. Has been a humbling realization, we’ve been overcomplicating this for years!”
Source: https://lnkd.in/p/g9j3bsgX
Join the discussion
Sign in with GitHub to leave a comment.