XNews

VikingRAG: Efficient RAG with Reduced Token Costs

VikingRAG: Efficient RAG with Reduced Token Costs

A new RAG method, VikingRAG, reduces token consumption to 5%–32% of state-of-the-art approaches while matching or exceeding accuracy. It addresses the high token cost of structured document retrieval. Key designs: hierarchical semantic storage (documents chunked with structural nodes; URIs like viking://Pasta/Carbonara/), evidence-gap-driven multi-round retrieval (agent decides if evidence sufficient, up to 15 rounds), and experience edges with adaptive upgrade (reuses past retrieval paths, upgrades to full multi-round only when needed). Tested on 6 datasets (0.24M–8.78M tokens) with DeepSeek-V4-Pro and other models. Basic VikingRAG uses 11.6%–51.9% tokens of SOTA; VikingRAG-E+ reduces to 5.1%–32.5% tokens. Traditional graph methods timed out on large datasets.

Key insights

Hierarchical Storage Eliminates Linear Directory Overhead

Instead of putting entire document structure in the prompt, VikingRAG stores all structural information in external, queryable objects. URIs encode ancestor-descendant relationships. Only visited fragments incur token cost, solving the linear growth problem of DeepRead.

Evidence-Gap-Driven Retrieval Matches SOTA Accuracy

Agent evaluates evidence after each tool call (Search, List, Grep, Read) and only continues if evidence is insufficient. Design separates coarse localization from fine validation, narrowing search space each step.

Experience Edges and Adaptive Upgrade Further Reduce Cost

VikingRAG-E stores edges from historical retrievals, allowing new queries to reuse paths (12%–33% token savings). VikingRAG-E+ upgrades to full multi-round only when constraint-based verification fails, adding 19%–50% savings.

Scalable Across Large Document Collections

VikingRAG's performance stays stable when documents increase from relevant-only to 991 total, unlike graph methods that timed out on large datasets. Insertion latency comparable to DeepRead/MoDora, faster than graph approaches.

Source posts