Noesis, a new Graph-RAG architecture described in a fresh arXiv paper, beats GraphRAG on a standard question-answering benchmark while using a much smaller on-premises model instead of GPT-4o.
The system targets three known problems with retrieval-augmented generation over knowledge graphs: static chunking that severs connections between sections of long documents, ingestion pipelines that do not scale, and multi-domain setups that either dilute retrieval accuracy in one giant knowledge base or need manual routing. Noesis pairs bidirectional graph traversal with an AIMD concurrency controller borrowed from TCP congestion control, delivering a claimed 23x ingestion speedup with zero out-of-memory failures. A companion technique called Moesis applies selective quantization to mixture-of-experts models, reportedly hitting 6.3x speedup on a 12 GB consumer GPU, while a routing layer called Mesh lets small on-premises models reason across separate knowledge bases at once. On HotpotQA (1,000 questions), the authors report 59.5 exact match and 74.7 F1, a 27.8-point exact-match gain over GraphRAG, using a 35B-parameter model for graph construction rather than GPT-4o.
The more interesting number isn't the benchmark score, it's the hardware. If a 35B on-premises model can out-score a GPT-4o-built graph pipeline, and a 12 GB consumer GPU can run mixture-of-experts inference at multi-x speedup, that changes who can realistically run Graph-RAG in-house instead of routing documents through a frontier-model API for indexing.
Worth remembering: these are the authors' own numbers on their own benchmark choice, not an independently replicated result, and HotpotQA is known to let systems shortcut genuine multi-hop reasoning.