Graph-free Multi-Hop RAG Framework MOTHRAG Released

Existing high-performance Multi-hop RAG systems such as GraphRAG, HippoRAG, and RAPTOR have a cost problem in that they must rebuild the knowledge graph through heavy LLM indexing whenever data changes.

The newly released MOTHRAG adopts a graph-free Dense Index and Query-time Orchestration approach, enabling immediate response to data updates through Embed-and-append alone, without retraining or rebuilding.

Key Benchmark Results (Accuracy/F1):

• HotpotQA: MOTHRAG 78.1 (GraphRAG 68.6, HippoRAG 75.5)
• 2WikiMultiHop: MOTHRAG 76.3 (GraphRAG 58.6, HippoRAG 71.0)
• MuSiQue: MOTHRAG 50.5 (GraphRAG 38.5, HippoRAG 48.6)

MOTHRAG can be run using only a general API without a GPU, and the cost per query is around $0.03. However, some performance degradation was observed on the MuSiQue benchmark, leaving this as a remaining challenge for improvement. This project has been released under the Apache-2.0 license.