VikingRAG Matches State-of-the-Art Structured-Document RAG Using 5.1-32.5% of the Tokens

arXiv·medium signalXBlueskyLinkedInCopy link

VikingRAG (arXiv 2609.11390) combines directory-aware structural retrieval with multi-round retrieval driven by evidence gaps. It stores successful multi-round traces as 'experience edges' that similar queries reuse, and escalates to the agentic loop only when one-round evidence falls short. The base system matched state-of-the-art accuracy on 11.6-51.9% of their tokens, and trace reuse plus adaptive escalation brought that to 5.1-32.5%. The idea of caching retrieval paths as well as documents transfers to any agentic RAG stack.