GLoRE: A Generalizable, Low-Resource–Oriented Retrieval-Augmented Generation Framework with Real-Time Deployment

Abstract

We present GLoRE—a Generalizable, Low-resource-oriented Retrieval-augmented gEneration framework that equips large language models with external knowledge for real-time, fact-grounded question answering across government and enterprise set- nectors, (ii) a three-stage retrieval stack enhanced by Contextual Retrieval—LLM-generated mini-summaries that raise chunk information density—and (iii) a novel Importance-aware Priority Weighting (IPW) mechanism that ensures authoritative passages surface even when purely semantic scores are mediocre. An optional conversation-memory layer sustains multi-turn coherence while keeping end-to-end inference under a two-second budget. Open-source reference code, empirical evaluations in mixed-language (low-resource) environments, and ablation studies on each module demonstrate GLoRE’s effectiveness and ease of adaptation, positioning it as a turnkey blueprint for trustworthy, priority-aware RAG deployments.

Keywords: Retrieval-Augmented Generation; Low-Resource NLP; Authority Ranking; Real-Time Systems

Date of Award	2025
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Kani CHEN (Supervisor)