A-RAG: Scaling Agentic Retrieval-Augmented Generation with Hierarchical Retrieval Tools

ABSTRACT

A-RAG proposes an agentic retrieval framework that exposes retrieval tools directly to large language models. Rather than treating retrieval as a static preprocessing step, the architecture allows models to dynamically choose between keyword search, semantic retrieval, and document reading. The design enables models to plan multi-step retrieval strategies that adapt to query complexity. This tool-based architecture allows RAG systems to scale with improving model reasoning capabilities. Experiments show improvements in both factual accuracy and retrieval efficiency.