VikingRAG is a directory-aware semantic data management system for retrieval-augmented generation over structured documents. It combines semantic and structural access and uses evidence-gap-driven multi-round retrieval to find missing facts. The base system matches state-of-the-art accuracy while using only 11.6% to 51.9% of the token cost of prior methods.
To cut multi-round token overhead, VikingRAG materializes successful retrieval traces as experience edges and reuses them for similar queries. An adaptive escalation strategy answers from one-round experience-augmented retrieval when possible, reserving multi-round agentic exploration for hard cases. With both enhancements, token cost drops to 5.1% to 32.5% of baselines while maintaining competitive accuracy and practical storage performance.
This is interesting because token cost, not just accuracy, is often the bottleneck for deploying RAG on large structured corpora. It also demonstrates a concrete way to close the loop between retrieval and reasoning without paying the full multi-round price every time.