Authorization-Aware Multi-Agent Retrieval-Augmented Generation for Secure Enterprise Knowledge Access

Retrieval-augmented generation (RAG) enables language models to answer enterprise questions using external documents; however semantic retrieval and multi-agent delegation create authorization risks when access checks occur only after search or answer generation. This study presents a literature-informed architecture for authorization-aware multi-agent RAG in which a policy decision point evaluates current user and resource attributes, whereas a policy enforcement point restricts the eligible candidate set before content reaches rerankers, agents, or a large language model. This design combines role-based and attribute-based access control, semantic embeddings, vector search, least-privilege delegation, provenance-aware auditing, and policy revalidation at protected access boundaries. A focused survey identifies risks including semantic overfetch, cross-domain synthesis leakage, stale permissions, and delegation escalation. The paper defines authorization-correctness conditions and proposes an evaluation protocol that compares authorization-first retrieval with a retrieve-then-filter baseline. The metrics include unauthorized-context exposure, authorized Recall@k, answer relevance, latency, revocation propagation time, delegation violations, and audit completeness. This study is a survey and design contribution; no benchmark results are claimed because the proposed prototype has not yet been empirically evaluated.