Cost‑Savvy RAG: Local Model Feeds a Flagship LLM in a Cascading Loop
An enterprise‑grade RAG pipeline stages a cheap, on‑prem LLM to pre‑filter queries, then hands only the top results to an expensive hosted model. The authors benchmark twenty open‑source models, showing up to 70% cost reduction without hurting answer quality, and they close the loop with automated validation to keep the cascade tuned.