A retrieval-augmented LLM framework for severity prediction of bug reports in cloud-based mobile applications

The effective maintenance of Mobile Applications (MAPPs) relies on the timely and accurate triage of bug reports, where severity labels determine the order of issue resolution. Manual severity assessment is often labor-intensive, error-prone, and inconsistent, highlighting the need for automated approaches that capture both semantic and contextual information. This study introduces RAG-GPT-SBR, a Retrieval-Augmented Generation framework that leverages a fine-tuned GPT-2 model for automated bug severity prediction in Cloud-based MAPPs. In addition to textual content, the framework incorporates non-textual metadata such as device type, operating system, crash signature, and bug frequency. Each new report is enriched with semantically related historical cases retrieved from a curated knowledge base using a hybrid BM25 and dense retrieval mechanism, forming structured input prompts for GPT-2. The fine-tuned model encodes this composite information and generates contextual embeddings, which are fed into a classification head to predict severity levels. Experiments conducted on a public Hugging Face bug report dataset demonstrate that RAG-GPT-SBR outperforms existing deep learning baselines, achieving gains of 7.98%, 8.20%, 7.90%, and 8.10% in accuracy, precision, recall, and F1-score, respectively. The results affirm that combining retrieval-augmented context with generative language models enhances reliability, scalability, and consistency in automated mobile application maintenance.