Lightweight Chunk Selection for Mobile RAG via Feature Alignment

This method reduces mobile RAG computation by treating chunk selection as an evidence-alignment problem using question hidden states, MoE routing signals, and chunk embeddings. It aims to select the most effective context without the overhead of full LLM-based compressors.