<![CDATA[The application of Large Language Models in the medical domain is often hampered by issues of hallucination and limited up-to-date knowledge. Retrieval-Augmented Generation offers a solution for connecting LLM with factual data, but the quality of RAG output is highly dependent on the accuracy of the information retrieval process. This study aims to analyze the effect of chunk size and embedding model variations on retrieval quality in a medical chatbot system at the Nusa Putra Farmedika General Clinic. The method used is a comparative experiment by testing three chunk size variations (256, 512, and 1024 tokens) and comparing the performance of two embedding models, BGE Small and MiniLM-L6. The evaluation was conducted automatically using the RAGAS framework, focusing on the Context Recall and Context Precision metrics. These findings were implemented into a medical chatbot prototype as a form of functional validation. The results showed an inverse relationship between chunk size and retrieval quality, with a chunk size of 512 tokens producing the best level of information granularity. The BGE Small model proved to be slightly superior to MiniLM-L6 in capturing the semantics of clinical text. The most optimal configuration was achieved by combining the BGE Small model with a chunk size of 512, which produced the highest average score of 0.59, Context Recall of 0.45, and Context Precision of 0.74. This study recommends this configuration as a technical standard for the development of medical chatbot as a foundational step to improve context relevance and mitigate the potential for hallucinations.]]>
Copyrights © 2026
