Pruning RAG context down to what the answer actually needs

The article describes how the company Kapa built a small LLM to prune 68% of the context from their retrieval API, keeping 96% of recall, and reducing the cost of a query by a third.