Re2GCC: Retrieve-Rerank-Generate Framework for Code Completion

Code pre-trained language models such as CodeGPT have demonstrated very promising results on the code completion task. The retrieval-augmented approach has further improved the accuracy of code completion. Previous retrieval-augmented approaches require a code-to-code retriever, but there is no standard dataset for retriever training, which usually requires manual construction of a training dataset containing semantically similar code pairs, and this is time-consuming and resource-intensive. We propose Re
²
GCC, a retrieve-rerank-generate framework for code completion. This framework obtains the most relevant code fragments to the unfinished code in the zero-shot setting by combining sparse retrieval and reranking techniques, which are then combined with an auto-regressive language model for generation. Compared to previous approaches, Re
²
GCC provides a low-cost solution for retrieval-augmented code completion tasks. Evaluating on the Python programming language, we demonstrate the superior performance of the Re
²
GCC framework on the code completion task.