Hybrid Retrieval and AI-Based Frameworks for Software Repository Understanding

This review gathers research on retrieval and AI-based software repository understanding across four main topics: lexical retrieval, dense representations, repository structure and evidence-grounded generation. The results show that adding BM25 to retrieval can help with both exact identifier searches and broader conceptual questions. Rank fusion and reranking can make the results more relevant. Organizing code chunks around AST boundaries keeps code units. Iterative retrieval and generation can add context from across a repository. Evidence-grounded generation lets developers compare AI explanations with sources but reliable attribution is still important. Together these methods give a base, for building repository assistants that provide relevant, context-aware and verifiable answers. Despite these results some problems still exist. Hybrid search systems can add time to process break down store indexes and work together. Current tests also only look at a programming languages while methods like changing how things are combined using groups of agents and making short summaries need more tests. Future studies should look into adding information without redoing everything combining searches based on what people are looking for and making smaller ways to show how parts of code are connected. Other important areas are learning from diagrams with code making it possible to use these tools on computers with smaller models and creating helpers that check and run the code they make. Solving these issues could make tools for code repositories better easier to use and more trustworthy, in software work.

[1]. H. Husain, H.-H. Wu, T. Gazit, M. Allamanis, and M. Brockschmidt, “CodeSearchNet challenge: Evaluating the state of semantic code search,” arXiv:1909.09436, 2019. [2]. Y. Wan et al., “Deep learning for code intelligence: Survey, benchmark and toolkit,” ACM Compute. Survey., vol. 56, no. 12, Art. 309, Oct. 2024, doi: 10.1145/3664597. [3]. P. Lewis et al., “Retrieval-augmented generation for knowledge-intensive NLP tasks,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020, pp. 9459–9474. [4].  Y. Gao et al., “Retrieval-augmented generation for large language models: A survey,” arXiv:2312.10997, 2023. [5]. M. Esposito, A. Janes, D. Taibi, and V. Lenarduzzi, “Generative AI in evidence-based software engineering: A white paper,” arXiv:2407.17440, 2024. [6]. M. Marcelino, M. Alves, B. Trinkenreich, B. Cartaxo, S. Soares, S. D. J. Barbosa, and M. Kalinowski, “Investigating the use of LLMs for evidence briefings generation in software engineering,” arXiv:2507.15828, 2025. [7]. X. Li et al., “CodeRetriever: Large-scale contrastive pre-training for code search,” in Proc. 2022 Conf. Empirical Methods Natural Lang. Process. (EMNLP), Dec. 2022, pp. 2898–2910. [8].  W. Gu, Z. Li, C. Gao, C. Wang, H. Zhang, Z. Xu, and M.R. Lyu, “CRaDLe: Deep code retrieval based on semantic dependency learning,” Neural Netw., vol. 141, pp. 385–394, Sept. 2021, doi: 10.1016/j.neunet.2021.04.019. [9]. W. Cheng, Y. Wu, and W. Hu, “Dataflow-guided retrieval augmentation for repository-level code completion,” in Proc. 62nd Annu. Meeting Assoc. Comput. Linguistics (ACL), Vol. 1: Long Papers, Aug. 2024, pp. 7957–7977. [10]. S. Zhang, Y. Ding, S. Lian, S. Song, and H. Li, “CodeRAG: Finding relevant and necessary knowledge for retrieval-augmented repository-level code completion,” in Proc. 2025 Conf. Empirical Methods Natural Lang. Process. (EMNLP), Nov. 2025, pp. 23278–23288.