Penerapan Synthetic Context Generation Menggunakan Large Language Model pada Sistem Question Answering Berbasis Retrieval-Augmented Generation untuk Domain Kesehatan Gizi

Abstract

The application of Natural Language Processing and Large Language Models in pediatric nutrition has become increasingly relevant due to the complexity of malnutrition management and the continued reliance on manual nutritional assessment processes. However, the development of Question Answering (QA) systems in this domain is challenged by the limited availability of structured narrative context within question–answer datasets. This study proposes the use of Synthetic Context Generation to construct a domain-specific knowledge base integrated into a Retrieval-Augmented Generation (RAG) architecture. Synthetic contexts are generated from patient diagnosis question-answer pairs using zeroshot and few-shot prompting strategies and indexed into a vector database as retrieval sources for the QA system. The quality of the synthetic contexts is evaluated using BLEU, ROUGE, and BERTScore metrics, while the QA system performance is assessed using the RAGAS framework. The results indicate a high level of semantic similarity between the synthetic contexts and reference data, reflected by consistently high BERTScore-F1 values above 0.8300, despite variations in BLEU and ROUGE due to paraphrasing and lexical differences. The RAG evaluation further demonstrates high context recall values exceeding 0.8700 and reaching up to 0.9600, indicating effective retrieval performance. Nevertheless, variations in faithfulness and answer correctness metrics reveal that the final answer quality is influenced by instruction restriction settings during the generation stage, highlighting the importance of instruction control to mitigate hallucination risks and enhance system safety in the nutrition health domain.