We taught a small LLM to throw away 68% of our RAG context