AutoRAGTuner: A Declarative Framework for Automatic Optimization of RAG Pipelines

AutoRAGTuner, presented by researchers Xintan Zeng, Yongchao Liu, Yice Luo, and Jiajun Zhen at EuroSys 2026, automates the notoriously manual process of optimizing Retrieval-Augmented Generation (RAG) pipelines. RAG enhances LLMs by grounding them in external knowledge, but its performance is highly sensitive to complex architecture designs and hyperparameters—requiring time-consuming manual tuning. AutoRAGTuner solves this with a declarative, configuration-driven approach that orchestrates the entire RAG lifecycle: construction, execution, evaluation, and optimization. Its modular architecture decouples pipeline stages through a component registration mechanism, while the Domain-Element Model (DEM) unifies heterogeneous data by representing objects as atomic elements with bidirectional pointers, supporting nodes, edges, and hyperedges. An integrated adaptive Bayesian optimization engine handles end-to-end hyperparameter tuning automatically.

Experimental results demonstrate AutoRAGTuner's generality across diverse RAG pipelines, from vanilla to graph-based, consistently outperforming default baselines. Notably, its declarative configuration language enables up to a 95% reduction in code churn for architectural adjustments, massively cutting engineering overhead. For developers and researchers building production RAG systems, this means faster iteration, less boilerplate, and more scalable, reusable architectures. By providing a systematically optimizable foundation, AutoRAGTuner paves the way for evolvable RAG systems that adapt without manual re-engineering—a significant practical advance for LLM-powered applications.