Rag

Create a tool to pre-calculate and cache KV states as 'Knowledge Packs', injecting them directly into the model for RAG tasks without the token overhead. This is a game-changer for latency and cost.