GRACE-RAG Enhances Closed-Domain RAG with Graph-Augmented Retrieval

Retrieval-Augmented Generation (RAG) systems are crucial for institutional question-answering, where responses must be grounded in authoritative documents. However, in complex domains with information spread across many documents, standard vector-only retrieval often yields fragmented evidence, burdening the LLM with extensive inference-time reasoning. GRACE-RAG addresses this by introducing a retrieval-governed, graph-augmented RAG architecture. This system shifts structural reasoning from the generative stage to a dedicated structured retrieval layer, resolving ambiguities offline. This design allows for deployment on lightweight, self-hosted models specifically calibrated for the institution's vocabulary. Experiments across various model capacities, including Mistral 24B, GPT OSS 120B, and Gemini 2.5 Flash, consistently showed GRACE-RAG improving response completeness, depth, and anticipatory coverage. Quality gains reached up to 20% with mid-scale models, indicating that the retrieval architecture's quality is more influential than model scale, leading to reduced computational and latency footprints without relying on proprietary solutions.