Citation-Enforced RAG Architecture for Decision Support

Abstract

Large language models (LLMs) are increasingly being deployed for institutional document synthesis and other reasoning-adjacent tasks. However, hallucinated content becomes an institutional risk when it is incorporated into briefs, policies, or reports. We present a local-first, corpus-restricted retrieval-augmented generation (RAG) system that enforces a system-level epistemic boundary through three mechanisms: (i) a strict offline corpus boundary, (ii) a traceability contract that requires factual claims to be supported by citations in a structured schema, and (iii) evidence-gated refusal when supporting evidence is weak or absent. We compare a single-pass RAG baseline with a multi-agent LangGraph workflow that separates the tasks of cited fact collection, reasoning over cited facts, and policy-oriented synthesis. Using a small, reproducible benchmark labeled as answerable or unanswerable, and considering two operating modes (analysis and policy), the multi-agent workflow reduces the proportion of unsupported claim units from 58.73% to 36.78% and improves refusal precision from 0.00% to 66.67%. However, it also exhibits citation index drift, which reduces citation accuracy from 87.50% to 62.50%. These results highlight both the value of architectural guardrails and the need for automated citation validation and repair in multi-step generation.