MOTHRAG matches GPU-bound multi-hop RAG systems via commodity APIs only
Julian Geymonat's MOTHRAG achieves an average F1 of 68.3 across HotpotQA, 2Wiki, and MuSiQue benchmarks — matching top multi-hop RAG systems like HippoRAG 2, CoRAG, and NeocorRAG — using only commodity API calls, no GPU, no fine-tuning, and no constrained decoding.
Score breakdown
MOTHRAG demonstrates that multi-hop RAG performance at the GPU-tuned state-of-the-art tier is achievable with commodity API calls alone, removing the GPU and fine-tuning infrastructure barriers that previously defined that performance level.
- 01MOTHRAG achieves an average F1 of 68.3 on HotpotQA, 2Wiki, and MuSiQue — within 0.7 points of the GPU-bound state of the art.
- 02It outperforms HippoRAG 2 (65.0 avg F1) and CoRAG (67.7 avg F1), and trails NeocorRAG (69.0 avg F1) by 0.7 points on average.
- 03MOTHRAG requires no GPU, no fine-tuning, no constrained decoding, and carries no non-commercial license restrictions.
Julian Geymonat published MOTHRAG, a multi-hop question-answering framework positioned against three of the strongest published systems in the space: HippoRAG 2 (offline graph + GPU, avg F1 65.0), CoRAG (trained retrieval, avg F1 67.7), and NeocorRAG (GPU constrained decoding, avg F1 69.0). MOTHRAG achieves an average F1 of 68.3 across HotpotQA (78.1), 2Wiki (76.3), and MuSiQue (50.5) using only commodity API calls — no GPU, no fine-tuning, no constrained decoding, and no non-commercial licenses. The post frames the goal not as beating these systems outright, but as reaching their performance tier without their infrastructure requirements.
The pipeline is fully modular: readers, embedders, and retrieval judges are all swappable without retraining, and are installed as optional extras.
The pipeline is fully modular: readers, embedders, and retrieval judges are all swappable without retraining, and are installed as optional extras. Supported components include `gemini`/`openai` for API-based readers and embedders, `sentence-transformers` for a local embedding fallback, `faiss` for vector stores over 100k–10M chunks, `retrieval` for classic BM25/graph features, and `prod` for the full stack. Every answer is proof-tree-structured, exposing each reasoning hop for inspection, and per-query outputs behind every benchmark table in the accompanying paper are released for independent verification. The paper is available on Zenodo and the code is released under Apache 2.0.
Key facts
- 01MOTHRAG achieves an average F1 of 68.3 on HotpotQA, 2Wiki, and MuSiQue — within 0.7 points of the GPU-bound state of the art.
- 02It outperforms HippoRAG 2 (65.0 avg F1) and CoRAG (67.7 avg F1), and trails NeocorRAG (69.0 avg F1) by 0.7 points on average.
- 03MOTHRAG requires no GPU, no fine-tuning, no constrained decoding, and carries no non-commercial license restrictions.
- 04Deployment is via `pip install mothrag` plus API keys; the pipeline is fully modular with swappable readers, embedders, and retrieval judges.
- 05A one-flag economy tier drops per-query cost from ~$0.032 to ~$0.018 at statistical parity on HotpotQA and 2Wiki.
- 06Every answer is proof-tree-structured, making each reasoning hop inspectable.
- 07Code is released under Apache 2.0; the paper is published on Zenodo.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 20, 2026 · 08:55 UTC. How this works →