المفاهيم الأساسية
Efficient code completion through selective retrieval with REPOFORMER.
الملخص
Recent advancements in retrieval-augmented generation have led to a new era in repository-level code completion. However, the indiscriminate use of retrieval poses challenges in efficiency and robustness. This paper introduces a selective RAG framework powered by REPOFORMER, improving performance while reducing latency. Extensive evaluations demonstrate the effectiveness of this approach across various benchmarks and programming languages.
الإحصائيات
Our framework consistently outperforms state-of-the-art prompting methods.
Selective retrieval strategy results in up to 70% inference speedup without compromising performance.
Performance improvements are observed across diverse benchmarks including RepoEval and CrossCodeEval.
اقتباسات
"Our findings suggest that the answer is predominantly negative, primarily for two reasons."
"REPOFORMER reflects three core principles: performance-oriented self-evaluation, robustness to retrieved contexts, and generalizability."
"Our approach's effectiveness in enhancing accuracy while significantly reducing latency showcases its potential in practical coding environments."