Research questionHow can soft context compression reduce RAG inference cost without inheriting the uncompressed system’s performance ceiling?Retrieved documents can make RAG inference expensive, and compressing them by imitating an uncompressed system may limit the compressed model to that system’s capabilities.