Recent advances in pre-trained Large Language Models (LLMs) have greatly improved automated code completion, but challenges such as logical errors, semantic misunderstandings, and hallucinated outputs persist, especially for complex or unseen code. Retrieval-Augmented Generation (RAG) alleviates these issues by retrieving external snippets as supplementary context, yet its single-view encoder often fails to capture the full range of code semantics. We propose Dual Multi-RAG, a novel framework that combines prompt-driven multi-retrieval with multi-view contrastive learning. In the retrieval stage, carefully designed prompts guide LLMs to elicit diverse semantic perspectives. In the selection stage, a contrastive learning mechanism identifies the most contextually relevant snippet for completion. This dual design enriches semantic coverage and improves reliability. Experiments demonstrate that Dual Multi-RAG consistently outperforms existing approaches, achieving 3.91% higher accuracy on the CCEval benchmark and 2.8% improvement on HumanEval-Infilling, confirming its effectiveness for robust and accurate code completion.
Shang et al. (Wed,) studied this question.