<![CDATA[Retrieval-augmented generation (RAG) is increasingly explored for clinical decision support, yet the contribution of the embedding model to retrieval quality on Indonesian medical documents has not been measured under controlled conditions. This study evaluates the effect of embedding model choice by comparing BAAI/bge-m3 and multilingual-e5-large on Indonesian clinical practice guidelines for hypertension (ICD-10 I10) and type 2 diabetes mellitus (ICD-10 E11). Two RAG pipelines were built to be identical except for the embedding component, using a FAISS index with cosine similarity, chunks of about 400 tokens with 15% overlap, and top-k fixed at 5. Ground truth consisted of 152 query–document pairs created through passage-to-query generation. Precision@5, Recall@5, MRR, and nDCG@5 were compared using paired Wilcoxon signed-rank tests, percentile bootstrap confidence intervals, rank-biserial effect sizes, and Holm-Bonferroni correction. BGE-M3 obtained higher mean scores on all four metrics, but only nDCG@5 differed significantly after correction (0.857 versus 0.806; Holm-adjusted p = 0.046). These results suggest that BGE-M3 is a better, or at least equivalent, choice, with its advantage clearest in ranking quality. This is a retrieval study and not a clinically validated decision support system.]]>
Copyrights © 2026
