Noesis: Graph-RAG breakthrough cuts retrieval costs 23x

Nicola Cogotti has introduced Noesis, a groundbreaking Graph-RAG (Retrieval-Augmented Generation over knowledge graphs) architecture that addresses critical limitations in traditional systems. The core innovation lies in its bidirectional graph traversal algorithm, which mimics human reading patterns with degrading memory to preserve semantic connections across long documents. This approach eliminates the static chunking problem that fragments context in conventional systems.

The system achieves remarkable performance gains through three additional components: an AIMD (Additive Increase Multiplicative Decrease) concurrency controller adapted from TCP congestion control that delivers 23x speedup without out-of-memory errors, Moesis—a domain-aware selective quantization technique for MoE models that provides 6.3x speedup on 12GB consumer GPUs, and Mesh, which enables cross-knowledge-base semantic routing with runtime structural discovery. On the HotpotQA benchmark with 1,000 questions, Noesis achieves 59.5 EM (Exact Match) and 74.7 F1 scores, outperforming GraphRAG by +27.8 EM while using a 35B on-premises model for graph construction instead of GPT-4o. The system also maintains 90% precision on long-range causal edges in a 193-page document verification test, where traditional chunk-independent extraction fails.