A user asked one simple question. By the time the RAG agent finished "figuring it out," it had quietly reprocessed the same information several times over — and nobody noticed until the bill did.
Turns out it wasn't the model. Wasn't the vector database either.
There are 5 specific places where RAG systems bleed cost like this — some hiding in techniques engineers use, thinking they're improving quality.
I broke down all five, and what I'd do instead.
Top comments (0)