Advising on academic regulations is difficult because sources are fragmented, frequently revised, and written in administrative prose. We present CRANE, a five-stage category-routed advising framework with two algorithms: CRANE Build for offline construction and CRANE Resolve for online answering. It unifies dense retrieval with graph-guided reasoning over a regulation-enriched graph. A lightweight classifier routes each query to a domain-specific subgraph and restricts vector search to the matching index partition. Graph evidence is gated by an edge-confidence threshold \(\tau \) and does not affect vector similarity. Vietnamese named-entity recognition and PhoBERT embeddings provide language-appropriate enrichment, and the generator receives a structured prompt with citations and version metadata. We evaluate CRANE on institutional question–answering (Q-A) using two resources: 13,684 labeled queries and 8,635 Q-A pairs verified by academic staff. Under protocol parity, CRANE attains an F1 Score of 99.03% with GPT and 98.69% with Mistral at \(\tau {=}0.6\) . It improves over a vector-only RAG baseline by 3.66 and 3.71 points while maintaining average end-to-end latency near 7 s per query. The domain router reaches 97.29% accuracy and 96.72% F1 Score on six-way domain classification. Ablation studies show that removing query routing, named-entity recognition, or budgeted \(\tau \) -thresholded graph expansion reduces F1 Score by 2.56–4.38 points. Sensitivity peaks at \(\tau {=}0.6\) . Scalability tests keep accuracy stable with near-linear latency as the vector index grows. Overall, these results indicate that category-routed retrieval with confidence-gated graph evidence can deliver verifiable regulation advising with high accuracy under a bounded latency budget.
