💡 The paper introduces Prometheus, a multi-agent system that resolves real-world issues in multilingual codebases using a unified knowledge graph. The problem addressed is that existing approaches to automated issue resolution are limited to Python-only issues and rely on pre-constructed containers, restricting their applicability to real-world and multi-language repositories.

To solve this problem, Prometheus transforms an entire code repository into a unified knowledge graph, encoding files, abstract syntax trees, and natural language text into a graph of typed nodes and five general edge types. This graph supports multiple programming languages and is persisted using Neo4j, enabling scalable and structured reasoning over large codebases.

The Prometheus system uses the DeepSeek-V3 model to resolve issues, and its effectiveness is demonstrated through experiments on SWE-bench Lite and SWE-bench Multilingual, where it resolves 28.67% and 13.7% of issues, respectively. The system achieves an average API cost of 0.23 and 0.38 per issue, and it resolves 10 unique issues not addressed by prior work.

The paper also shows that Prometheus is the first system to demonstrate effectiveness across seven programming languages and can resolve real-world GitHub issues in the LangChain and OpenHands repositories. The Prometheus system has been open-sourced, making it available for further development and use. Overall, the paper presents a significant contribution to the field of automated issue resolution in multilingual codebases, providing a unified and scalable approach to resolving real-world issues.