TopPaper
Today's Paper

GRACE-RAG: Governed Retrieval Architecture for Canonical Evidence Synthesis, Enabling Lightweight Deployment in Closed-Domain Institutional Settings

Asit Desai, Prashant Devadiga, Aman Kumar • arXiv (cs.IR)

GRACE-RAG introduces a retrieval-governed, graph-augmented architecture that resolves structural entity ambiguity offline prior to LLM generation. By restricting the generative model purely to evidence synthesis over dual embedding surfaces, it boosts response quality by up to 20% on mid-scale models. This framework enables self-hosted 24B parameter models to match the structural quality of proprietary frontier systems in complex institutional question answering.

View Full Abstract DOI: 10.48550/arXiv.2607.00013 Share

Keywords

Retrieval-Augmented Generation Knowledge Graphs Entity Canonicalization Dual Embeddings Institutional QA