AQACO: Adaptive Query-Aware Chunking Optimization for Retrieval-Augmented Generation Systems

Tharakesavulu Vangalapat, Lohith Kumar Deshpande

Abstract: Background: Retrieval-Augmented Generation (RAG) systems have revolutionized knowledge-intensive natural language processing tasks, yet their performance is fundamentally constrained by static document chunking strategies that ignore query characteristics and domain-specific requirements. Methods: We introduce AQACO (Adaptive Query-Aware Chunking Optimization), a novel framework that dynamically optimizes chunking parameters through multi-objective Bayesian optimization combined with reinforcement learning. Our approach analyzes query patterns to predict optimal chunking strategies, considering retrieval quality, context completeness, and computational efficiency simultaneously. We evaluated AQACO on six public datasets across four domains (MS MARCO, Natural Questions, HotpotQA, FEVER, SciFact, and FiQA-2018). Results: AQACO achieves substantial improvements: 24.3% higher NDCG@5, 28.7% reduction in answer fragmentation, and 19.4% lower processing latency compared to state-of-the-art static chunking methods. Conclusions: Our query-aware optimization paradigm establishes new benchmarks for adaptive document processing in RAG systems, with open-source implementation and reproducible experiments available for the research community.

Keywords: Retrieval-Augmented Generation, Document Chunking, Adaptive Systems, Bayesian Optimization, Information Retrieval, Natural Language Processing

How to Cite?: Tharakesavulu Vangalapat, Lohith Kumar Deshpande, "AQACO: Adaptive Query-Aware Chunking Optimization for Retrieval-Augmented Generation Systems", Volume 14 Issue 3, March 2025, International Journal of Science and Research (IJSR), Pages: 1714-1722, https://www.ijsr.net/getabstract.php?paperid=SR25329182150, DOI: https://dx.doi.org/10.21275/SR25329182150