TRUST-Campus: Evidence-Aware RAG for Reliable University Information Access

TRUST-Campus is an evidence-aware retrieval-augmented generation framework designed to improve the reliability of university information access. Rather than treating retrieval relevance alone as sufficient grounds for answering, the system evaluates whether retrieved evidence is adequate to support a response and abstains when that evidence is insufficient. The framework combines dense and BM25 retrieval, cross-encoder reranking, semantic evidence-sufficiency assessment, and selective answer-or-abstain decisions. It is evaluated on a frozen 50-question benchmark spanning paraphrase, compositional, negation, personal-context, unsupported, future, and false-premise queries. Results show that the Evidence-Aware V4 system improves answer/abstain decision accuracy and reduces false rejections relative to dense, hybrid, and reranked RAG baselines, while exposing remaining trade-offs between coverage and false acceptance. This work studies a broader question in trustworthy AI: when should an intelligent system act on available evidence, and when should it decline to answer because the available evidence is not sufficient? Status: Preprint / unpublished research manuscript.