RABTR: Retrieval-Augmented Bug Triage With Reasoning

Abstract

Automatic bug triage aims to recommend appropriate developers for newly reported issues. While prior work has focused on improving recommendation accuracy, little attention has been given to supporting developers with explanations for these recommendations. In this paper, we present a retrieval-augmented framework for bug triage that integrates both predictions and explanations. We propose retrieval-based automatic bug triage (RABT), a semantic retrieval method that identifies similar past issues using vector-based embeddings, and retrieval-augmented bug triage with reasoning (RABTR), a reasoning method that generates natural language explanations grounded in retrieved evidence. To assess the practical utility and robustness of our approach, we conduct experiments on six real-world datasets. We further perform a time-split evaluation to reflect real-world maintenance scenarios, where RABT consistently outperforms state-of-the-art methods, including TF-IDF, BM25, and DBRNN-A, with an improvement of up to 26.10 percentage points in top-10 accuracy. To evaluate the quality of the generated explanations, we introduce a checklist-based evaluation method combining GPT-4.1 and human annotation. The results show strong alignment between model and human evaluations on key dimensions such as accuracy, precision, and conciseness. The ablation analysis also reveals that explanation quality is highly sensitive to retrieval input, reinforcing the importance of semantic retrieval in reasoning-augmented systems. Our findings highlight the potential of integrating semantic retrieval and large language model-based reasoning to improve both the transparency and effectiveness of automatic bug triage.