Repowise: A Hybrid RAG-Based Conversational System for Software Repository Analysis
8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications, ICHORA 2026, Ankara, Turkey, 21 - 23 May 2026, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/ichora69329.2026.11537220
- City: Ankara
- Country: Turkey
- Keywords: Code Question Answering, Knowledge Graphs, Large Language Models, LLM-as-Judge, Retrieval-Augmented Generation, Software Repository Mining
- Hacettepe University Affiliated: Yes
Abstract
Software repositories are growing rapidly, making it increasingly difficult for developers to understand codebases, track dependencies, and resolve issues efficiently. Traditional semantic search methods often fail to capture the complex architectural relationships within a project. This paper presents RepoWise, a repository analysis system that combines a hybrid Retrieval-Augmented Generation (RAG) pipeline with an intentbased routing mechanism. By integrating a vector database (ChromaDB) for semantic code search and a knowledge graph (Neo4j) for mapping repository structures and commit histories, RepoWise routes queries to the most appropriate data source to answer developer questions accurately. We evaluate the system on a custom benchmark of 60 questions across five categories (Metadata, Semantic, Structural, Historical, and Hybrid) derived from two real-world open-source repositories-psf/requests and pallets/flask. Using an LLM-as-Judge evaluation framework with 5-round majority voting, RepoWise achieves an overall accuracy of 78.3%(47/60), with perfect scores on Metadata (100%) and strong performance on Semantic (94.1%) categories. These results demonstrate the effectiveness of hybrid graph-vector retrieval for code repository question answering, while revealing structural and historical query types as the primary areas for future improvement.