Researchers have introduced Noesis, a novel Graph-RAG architecture designed to overcome limitations in grounding large language models with domain-specific knowledge. Noesis employs four key algorithms: Bidirectional Graph Traversal for simulating human reading, an AIMD Concurrency Controller for adaptive scaling that achieved a 23x speedup, Moesis for domain-aware quantization offering a 6.3x speedup on consumer GPUs, and Mesh for cross-knowledge-base semantic routing. This system demonstrated significant improvements on the HotpotQA benchmark, achieving 59.5 EM / 74.7 F1, a 27.8 EM increase over existing Graph-RAG methods, while utilizing a smaller on-premises model for graph construction instead of GPT-4o. AI
IMPACT Enhances grounding of LLMs in domain-specific corpora, potentially improving accuracy and efficiency in knowledge-intensive tasks.
RANK_REASON The cluster describes a new research paper detailing a novel architecture for Graph-RAG.
Read on arXiv cs.IR (Information Retrieval) →
- AIMD Concurrency Controller
- GPT-4o
- Graph-RAG
- HotpotQA
- knowledge graphs
- large language models
- Mesh
- MoE models
- Nicola Cogotti
- Noesis
- arXiv
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
