A-RAG Scales Agentic RAG with Hierarchical Retrieval Interfaces

Summary

The paper introduces A-RAG, an agentic retrieval-augmented generation framework designed to let frontier language models participate directly in retrieval decisions. The authors argue that conventional RAG either retrieves passages in one step and concatenates them into the prompt or follows a workflow defined in advance, limiting how effectively retrieval can benefit from model improvements. A-RAG exposes three hierarchical retrieval interfaces to the model: keyword search, semantic search, and chunk read. An agent can use these tools repeatedly and at different granularities, adapting its search process to the task instead of following a fixed retrieval procedure. Experiments across multiple open-domain question-answering benchmarks show that A-RAG consistently outperforms existing approaches while using a comparable or smaller number of retrieved tokens. The authors also systematically examine how performance scales with model size and test-time compute. The paper reports that the framework can leverage stronger models and additional inference-time computation, although the supplied abstract does not provide detailed benchmark scores or identify the evaluated models. The authors state that the code and evaluation suite are available for future research.