AgenticRAG-Plan Improving Retrieval-Augmented Generation with Multi-Agent Reasoning

Abstract

This paper presents AgenticRAG-Plan, a multi-agent framework for Retrieval-Augmented Generation (RAG) that combines specialized AI agents within an adaptive retrieval and reasoning pipeline. Unlike conventional RAG systems that follow a fixed retrieve-then-generate workflow, the proposed framework introduces four dedicated agents: a Planner Agent for query decomposition, a Retriever Agent for document retrieval, a Generator Agent for response generation, and a Verifier Agent for evaluating factual consistency and guiding iterative refinement. By coordinating these agents through a feedback-driven workflow, the framework supports adaptive retrieval, continuous refinement, and dynamic decision-making that extend beyond the capabilities of traditional RAG pipelines.