This paper introduces MetaRAG, an Adaptive Retrieval-Augmented Generation system designed to address multiple-choice questions in the field of Psychology and Psychiatry in Vietnamese. MetaRAG leverages adaptive retrieval mechanisms to extract relevant information based on user input and integrates this data through multi-threaded analysis to deliver optimal answers. The study evaluates the performance of large language models fine-tuned with real-world and synthetic data, highlighting the impact of data augmentation on system accuracy and robustness. Furthermore, it examines the influence of embedding size variations on RAG model performance. The results demonstrate MetaRAG’s effectiveness in enhancing Vietnamese mental health question-answering systems, providing valuable insights into applying LLMs and RAG pipelines in domain-specific contexts.