Source code for our EMNLP 2025 paper: "CodeRAG: Finding Relevant and Necessary Knowledge for Retrieval-Augmented Repository-Level Code Completion" [arXiv].

uv sync

source .venv/bin/activate

Before running scripts, download benchmarks (recceval and cceval) and edit the configuration file:

config/config.toml

Then execute the Python scripts sequentially:

python scripts/build_query.py

• Generates query strings from the benchmark dataset.

python scripts/retrieve.py

• Retrieves top-k relevant code blocks using the configured retriever.

python scripts/rerank.py

• Reranks retrieved code blocks based on their estimated importance.

python scripts/build_prompt.py

• Constructs prompts from retrieved code blocks for the code completion generator.

python scripts/inference.py

• Feeds prompts to the generator model.
• You can replace this step with your own inference code.
  Input: JSON file containing an array of strings
  Output: JSON file containing an array of generated completions.

python scripts/evaluation.py

• Evaluates code completion performance using inference results.

If you find this work helpful, please consider citing our paper:

@inproceedings{coderag2025, title={CodeRAG: Finding Relevant and Necessary Knowledge for Retrieval-Augmented Repository-Level Code Completion}, author={Sheng Zhang, Yifan Ding, Shuquan Lian, Shun Song, Hui Li}, booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)}, year={2025} }

For questions, please open an issue or contact dingyf@stu.xmu.edu.cn.