Cutting RAG inference costs 6x starts with deciding what never reaches the LLM

Here is what changes when you cannot afford to be probabilistic about everything, and how a cascade architecture solves it.