REBot: From RAG to CatRAG with Semantic Enrichment and Graph Routing

Researchers introduced REBot, an LLM-powered chatbot that uses CatRAG, a hybrid retrieval-augmented generation framework combining dense retrieval with graph-based reasoning, to provide accurate academic regulation advising. The system achieved 98.89% F1 score on classification and question-answering tasks and demonstrates how specialized domain knowledge graphs can enhance AI advisory systems.

REBot represents a meaningful advancement in domain-specific AI applications by addressing a critical gap in academic institution support systems. Rather than relying solely on general-purpose language models, the researchers engineered a specialized architecture that combines multiple retrieval strategies to ensure both factual accuracy and contextual reasoning. This hybrid approach reflects growing recognition that RAG systems alone may lack sufficient depth for complex regulatory interpretation tasks.

The development of CatRAG stems from broader trends in AI where practitioners recognize that retrieval quality directly impacts LLM output reliability. By integrating semantic enrichment within a hierarchical, category-labeled knowledge graph, the system provides structured domain understanding alongside traditional dense retrieval. The lightweight intent classifier further optimizes performance by routing queries intelligently rather than applying uniform processing across all requests.

For educational institutions and advisory service providers, REBot suggests a viable path toward reducing operational burden while maintaining compliance accuracy. The 98.89% F1 score substantially exceeds typical chatbot performance on specialized regulatory tasks, implying genuine practical utility. This could reshape how universities handle policy interpretation at scale, enabling consistent student guidance without proportionally increasing advising staff.

Looking forward, the CatRAG framework's architecture may influence how other domain-specific AI systems approach knowledge representation. Educational institutions considering similar implementations should evaluate whether their regulatory complexity justifies custom knowledge graph construction. The success here demonstrates that task-specific engineering outperforms generic LLM deployment for high-stakes advisory functions.

• →CatRAG hybrid framework combines dense retrieval and graph reasoning for superior regulatory domain performance.
• →REBot achieved 98.89% F1 score, substantially outperforming general-purpose chatbots on academic policy interpretation.
• →Semantic enrichment and category-labeled knowledge graphs enable factual accuracy critical for compliance advising.
• →Intent routing classifier optimizes which retrieval modules handle specific query types for efficiency.
• →Educational institutions can use this architecture model to build specialized AI advisory systems reducing manual workload.