AlignCoder: Aligning Retrieval with Target Intent for Repository-Level Code Completion

AlignCoder is a new repository-level code completion framework designed to enhance the effectiveness of code large language models (LLMs) by addressing the misalignment issues in retrieval-augmented generation (RAG) approaches. It introduces a query enhancement mechanism and a reinforcement learning-based retriever training method, resulting in an 18.1% improvement in exact match (EM) scores on the CrossCodeEval benchmark when evaluated across five different backbone code LLMs. This advancement is significant for practitioners as it enhances the contextual understanding of repository-specific code, improving the accuracy and reliability of code completion tasks.