Answers are grounded in this paper and cite their sources.
Understanding an unfamiliar software repository is difficult because modern codebases contain interconnected files, modules, dependencies, and configuration artifacts. Traditional keyword-based search provides limited contextual understanding, while directly processing large repositories using Large Language Models (LLMs) introduces context and scalability challenges. This paper presents RepoXray, a retrieval-augmented framework for repository-level software understanding. The system combines file-level LLM analysis, semantic embeddings, vector retrieval, and Map-Reduce-based synthesis to generate repositorylevel architecture and security-oriented reports. It also provides an AI Chat Console that retrieves relevant repository information to answer natural-language queries, along with an interactive File Explorer for inspecting source code and AI-generated file-level analysis. RepoXray was evaluated using 20 public GitHub repositories across four programming languages. It achieved 84.91% architecture fact coverage and 84.75% repository-aware question-answering accuracy. The preliminary combined score was 84.83%. Security performance could not be quantitatively validated because the evaluation dataset contained no positive ground-truth vulnerability instances. These results demonstrate the potential of RepoXray for improving repository understanding while highlighting the need for further evaluation.
No takes yet. Share an insight, caveat, or question.
Gohain et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: