This project investigates how AI—particularly large language models (LLMs)—can assist software developers by generating and adapting code based on contextual information. While modern AI coding assistants can produce syntactically correct code, they often struggle to account for the broader context of a project, such as existing classes, dependencies, APIs, and architectural constraints.

The goal of this project is to design and evaluate a system that improves code generation by incorporating different types of context, including: local context (method or function description), structural context (file, class, and package structure), dependency context (APIs, libraries, imports), semantic context (functional requirements and intent).

The student will implement a context-aware pipeline, potentially using retrieval-augmented generation (RAG), where relevant code snippets or project artifacts are retrieved and provided to the model during generation. The project will compare context-aware generation with standard prompt-based approaches to assess improvements in correctness, integration, and usability.

Interesting research questions that can be explored in a scope of this project:

• How does different types of context affect code generation quality?
• Does more context always help, or can it overload the model?
• Which context (API, structure, semantics) is most useful?
• Can retrieval improve reuse of existing code components?

References

1. Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J. and Awadallah, A.H., 2024, August. Autogen: Enabling next-gen LLM applications via multi-agent conversations. In First conference on language modeling. https://www.microsoft.com/en-us/research/wp-content/uploads/2023/08/LLM_agent.pdf
2. Tao, Y., Qin, Y. and Liu, Y., 2025. Retrieval-augmented code generation: A survey with focus on repository-level approaches. arXiv preprint arXiv:2510.04905. https://arxiv.org/abs/2510.04905