AutoRAGBench: A Benchmark for LLM Agents Automating RAG System Development

Hang Zhang, Yuzhi Pan, Ziyao Zhong, Chaokun Wang and colleagues at Tsinghua University and Sina Weibo introduce AutoRAGBench, which measures how well coding agents build retrieval-augmented generation systems from scratch and improve them through iteration.
Ask this paper
Setup. Nine RAG development tasks with no starter code and no prescribed strategy, under a six-hour budget on a single H800 GPU. Agents analyze the data, implement pipelines and refine them from experimental feedback.
Agents beat the human baseline. GPT-6.1 Sol with Codex CLI averages 56.1, which is 9.7 points above Reference RAG and 11.7 points above the composite Human-SOTA baseline of 44.4.
Improving human systems. Starting from a human-designed system on seven tasks, the same agent raises the score by 16.6 points on average.
Trajectory analysis. Nine agent configurations are compared on both final system quality and how their development process unfolds, including open models run in Claude Code and OpenCode.
Abstract
Retrieval-augmented generation (RAG) is a well-established paradigm for augmenting large language models (LLMs) with information retrieved from external corpora. However, developing effective RAG systems requires substantial engineering effort and iterative experimentation, as effective design choices vary across tasks and domains. With rapid advances in LLM agent capabilities, we ask how well these agents can autonomously build RAG systems and how effectively they can improve them through iteration. In this paper, we propose AutoRAGBench, the first benchmark to evaluate how well LLM agents develop RAG systems from scratch in an open code space. It comprises nine RAG system development tasks. For each task, agents build RAG systems from scratch without starter code or prescribed strategies, under a six-hour budget on a single H800 GPU. They analyze task data, implement RAG pipelines, and iteratively refine their designs based on experimental feedback. We evaluate nine agent configurations and analyze both the effectiveness of the resulting RAG systems and their development trajectories. GPT-6.1 Sol with Codex CLI achieves the highest average score of 56.1. The composite Human-SOTA baseline averages 44.4. These results show that LLM agents can automate effective RAG system development. When starting from a human-designed system on each of seven tasks, GPT-6.1 Sol with Codex CLI improves the score by 16.6 points on average.