How would you optimize a RAG system that is currently returning noise, which increases context entropy and causes attention decay in the token sequence?