MOTHRAG beats GraphRAG, HippoRAG 2, and RAPTOR on multi-hop benchmarks

A graph-free dense index system named MOTHRAG outperformed graph-based RAG systems on HotpotQA, 2WikiMultiHopQA, and MuSiQue benchmarks. The authors argue that for frequently changing data, the accuracy gains from knowledge graphs are outweighed by the constant cost of rebuilding them.

• On HotpotQA, MOTHRAG scored 78.1 compared to GraphRAG's 68.6, HippoRAG 2's 75.5, and RAPTOR's 69.5.
• On 2WikiMultiHopQA, MOTHRAG achieved 76.3, surpassing GraphRAG (58.6), HippoRAG 2 (71.0), and RAPTOR (52.1).
• On MuSiQue, MOTHRAG reached 50.5, beating GraphRAG (38.5), HippoRAG 2 (48.6), and RAPTOR (28.9).
• Updates require only embedding and appending without retraining or rebuilding, costing approximately $0.03 per query on commodity APIs.

The authors conclude that for multi-hop retrieval over dynamic data, graph-free indexing with strong query-time orchestration offers a more efficient alternative to the overhead of maintaining knowledge graphs.