Retrieval-Augmented Generation (RAG) has emerged as an effective solution to address limitations of Large Language Model (LLM) in accessing external knowledge sources, incorporating up-to-date information, and reducing hallucinations. These advantages are especially critical in domains that demand continuous updates and extremely high accuracy, such as the legal field. This study explores advanced retrieval strategies to enhance Vietnamese legal RAG systems. Through comprehensive experiments on Vietnamese and English datasets, we reveal that tuning the unigram weight parameter \(\boldsymbol{\lambda }\) in the combined BM25 (comBM25) model significantly improves key retrieval metrics. We further demonstrate that hybrid retrieval methods, integrating both Sparse and Dense Retrieval signals, consistently outperform individual approaches. Our experiments, conducted on Vietnamese legal datasets, show promising improvements. In addition to enhancements of the retrieval algorithm, we also propose effective methods for chunking legal documents to further improve retrieval relevance. These findings aim to support the development of more accurate and reliable RAG systems for the Vietnamese legal domain.