An Agentic RAG Architecture for Intelligent Knowledge Retrieval Systems

Abstract

This demo study presents an agentic retrieval-augmented generation architecture for institutional knowledge retrieval. The proposed design separates query planning, evidence retrieval, source evaluation, response synthesis, and citation checking into observable components coordinated by a constrained agent loop. A synthetic document collection and fictional user tasks are used to assess retrieval coverage, answer grounding, latency, and failure recovery. The simulated evaluation suggests that explicit planning and evidence verification can improve traceability when compared with a single-pass retrieval pipeline, although additional orchestration increases computational cost and operational complexity. The architecture therefore includes bounded tool use, confidence thresholds, audit records, and a fallback path for human review. Particular attention is given to document permissions and the prevention of unsupported claims. All documents, measurements, participant identities, and reported findings are fictional. This article is original demo content created solely for user-interface, search, galley, and publication-workflow testing in OJS.