GraphRAG with Sub-30B Models Excels at Cross-Document Multi-Hop Reasoning in Marathi
kmeanskaran · x · 2026-08-18
The author benchmarked GraphRAG against VectorRAG on Marathi government scheme documents, with deliberately mixed evaluation data. Key finding: GraphRAG works exceptionally well for cross-document multi-hop reasoning, even with sub-30B models.
- Setup: 29 Marathi documents through identical text and the same encoder (nomic-embed-text-v2-moe). GraphRAG extracted SVO triples via LLM into Neo4j (Entity-[:REL]->Entity); VectorRAG chunked into ChromaDB with top-k=5.
- Eval set: 36 questions—20 cross-document multi-hop, 8 unanswerable traps, 6 multi-fact single-doc, 2 direct-fact; 28 required cross-document reasoning, each annotated with bridgeentity and bridgepath.
- Queries covered one-hop (direct facts) and two-hop (A → hub → B) paths, showing small models plus graph structure can handle complex reasoning.
Related event: Study finds GraphRAG outperforms vector RAG in multi-hop reasoning(2 posts)→
More from Research
- UCSD Researchers Propose Dataset Watermarking Detectable via Black-Box API Access — yuxiangw_cs · 2026-10-16
- EmbodiedEvo tops RoboDojo as agents inherit reusable skills and lessons — Aiden_Tech_Ai · 2026-10-16
- CHI reviewer laments: papers' Discussion sections now cite nothing — IanArawjo · 2026-10-16
- Casper: Understanding 90% of neurons ≠ understanding the network — StephenLCasper · 2026-10-16
- NVIDIA Launches cuPotential in CUDA-X to Bring AI Agents to Semiconductor Materials Simulation — tomaszbednarz · 2026-10-16
- Agentic Search Systems Are Rewarded Only for Final Answers, Raising Faithfulness Concerns — burkov · 2026-10-16