DOTRAG redefines GraphRAG with retrieval-time reasoning for multi-hop queries
New training-free framework achieves SOTA on MetaQA and UltraDomain by reasoning along paths.
Deep Dive
DOTRAG is a training-free GraphRAG framework that reframes retrieval as a reasoning process over paths. It uses Division of Thought (DOT) to decompose retrieval into localized search spaces, pruning irrelevant regions and iteratively discovering relational paths. DOTRAG achieves state-of-the-art performance on MetaQA and UltraDomain, particularly excelling at complex multi-hop tasks.
Key Points
- Training-free GraphRAG framework that reformulates retrieval as reasoning over paths
- Uses Division of Thought (DOT) to decompose retrieval into query-specific localized search spaces
- Achieves SOTA on MetaQA and UltraDomain benchmarks, especially on multi-hop tasks
Why It Matters
Makes multi-hop knowledge retrieval far more accurate without expensive fine-tuning—a big win for enterprise RAG systems.