01933nas a2200109   4500000000100000008004100001100001500042700001900057700002000076245009100096520163600187       2026                            d1 aMert Nakip1 aRafał Gibała1 aSławomir Nowak00aOptimizing Information Retrieval in RAG Systems via Proactive Hyperparameter Selection3 a<p>Information retrieval is a crucial operation for Retrieval-<br>Augmented Generation (RAG) systems, providing substantial gains in<br>factual accuracy and operational eciency. However, the performance of<br>a RAG system is highly sensitive to the selection of hyperparameters, optimization of which is computationally expensive and time-consuming. In<br>this paper, we propose a novel framework, called Proactive Hyperparameter<br>Estimation and Selection (ProHES), which introduces a paradigm<br>shift from reactive, brute-force search to an ecient, analytical approach.<br>The core novelty of the ProHES framework is its two-stage methodology.<br>First, it analytically estimates the most suitable similarity metric<br>for a given embedding and dataset by quantifying intrinsic embedding<br>properties, i.e. sparsity, magnitude meaning, feature independence, and<br>value spread. Then, it systematically selects the optimal number of retrieved<br>documents based on user preferences and system requirements.<br>We evaluate the performance of the ProHES framework across six different<br>embedding functions on the MS Marco and TriviaQA datasets.<br>Our results demonstrate that ProHES provides a near-optimal hyperparameter<br>conguration with 22- to 86-fold reduction in computation<br>time compared to exhaustive search, achieving an average of 94% of the<br>best possible performance. The proposed framework represents a significant<br>advancement by making hyperparameter tuning for RAG systems<br>more ecient and systematic, thereby paving the way for more accurate,<br>robust and scalable real-world applications.</p>