VikingRAG: Accurate and Token-efficient Retrieval-augmented Generation over Structured Documents

VikingRAG implements a directory-aware semantic data management system to minimize structural-context tokens in RAG workflows. The method uses adaptive escalation to switch between one-round experience-augmented retrieval and agentic multi-round retrieval, while materializing retrieval traces as reusable experience edges to reduce repeated token overhead.