TaxGraphRAG: Ontology-Constrained Retrieval and Generation for Auditable Tax Advisory Systems

The exponential growth of complex tax legislation and unstructured regulatory documents necessitates reliable automated tax information services. While standard Retrieval-Augmented Generation (RAG) systems are currently used, they struggle with semantic ambiguity, temporal constraints, logical consistency, and source auditability in high-stakes regulatory domains. In this paper, we propose TaxGraphRAG, a three-tier ontology-constrained hybrid graph and retrieval-augmented generation framework designed to deliver trustworthy tax information. The architecture comprises three core modules: an Ontology-Constrained Graph Module (OCGM) that constructs a multi-layer knowledge graph mapping tax entities to strict domain semantics; a Hybrid Retrieval Engine (HRE) that combines vector similarity with constrained graph-based traversal; and a Constrained Generation Module (CGM) that enforces ontological compliance during decoding. Experimental evaluation demonstrates that TaxGraphRAG significantly outperforms standard RAG baselines, achieving an Ontology Compliance Score (OCS) of 0.847 compared to 0.287. Furthermore, the framework reduces hallucination rates by 67.4%, improves regulatory compliance from 67.3% to 94.8%, and increases factual accuracy to 89.6%. By maintaining explicit traceability to authoritative sources and enforcing structural rules, TaxGraphRAG establishes a highly accurate, interpretable, and logically consistent approach for automated tax advisory systems.