Citation-Enforced RAG for Fiscal Document Intelligence: Cited, Explainable Knowledge Retrieval in Tax Compliance

Citation enforcement reduces hallucination in high-stakes domains: The framework enforces citations during generation and supports "abstention when evidence is insufficient," meaning the system refuses to answer rather than guess—a critical safeguard for tax compliance where incorrect guidance could have serious legal and financial consequences.

Explainability is essential for regulatory adoption: The paper demonstrates that preserving "page-level provenance" and providing "analyst-usable explanations" are necessary for tax authorities and auditors to trust AI recommendations, since decisions must be traceable and defensible under scrutiny.

Multimodal document processing handles real-world complexity: Tax documents include forms, instructions, and guidance in mixed formats; the framework's multimodal approach addresses the reality that fiscal information exists across "unstructured and semi-structured" sources, not just clean text.

Conservative AI behavior protects against costly errors: By prioritizing "conservative behaviour" and refusing to answer uncertain questions rather than providing plausible-sounding but potentially incorrect information, the system acknowledges that in compliance contexts, silence is safer than speculation.

Responsible AI requires domain-specific design: Rather than applying generic AI solutions, this work shows that trustworthy systems for regulated industries need purpose-built features like citation enforcement and auditability—suggesting that responsible AI implementation cannot be one-size-fits-all.