Abstract: Poor digital infrastructure and a lack of local language resources result in Sub-Saharan Africa having limited access to specialized medical information. Swahili, spoken by over 100 million people, is crucial for the democratization of digital healthcare. The lack of an extensive Swahili medical corpus complicates the development of credible and culturally tailored language models. Large Language Models (LLMs) promise to be a boon for medical chatbots, providing accessible, context-aware, and language-appropriate health information. However, fine-tuning techniques have constraints, including diminished factual robustness and a larger risk of medical hallucinations. The Swahili-speaking medical chatbot Lueji, initially based on a fine-tuned model, is extended with a Retrieval-Augmented Generation (RAG) architecture to address these limitations. A FAISS-based semantic retriever utilizing a Swahili-translated Huatuo-26M Chinese medical corpus and the UlizaLlama model, tailored for African languages, is employed in the proposed system. This hybrid methodology improves factual dependability and contextual accuracy while retaining generative fine-tuning-achieved verbal fluency. The BLEU, ROUGE, and GLEU metrics were used to quantify lexical consistency, structural similarity, and text quality. The fine-tuned model outperforms the RAG-based version on typical text similarity measures (BLEU-1 = 0.2217, ROUGE-1 = 0.3168, GLEU = 0.1077 vs. 0.1912, 0.2916, 0.0918). A qualitative investigation reveals that the RAG architecture significantly reduces hallucinations and enhances clinical factuality. Low-resource environments present a fundamental trade-off between linguistic fluency and factual accuracy. The study gives fresh empirical insights into the balance between fine-tuning and retrieval augmentation for low-resource medical LLMs. It lays the groundwork for reliable, hybrid, and culturally inclusive African medical chatbots.

Keywords: Swahili Chatbot, Retrieval-Augmented Generation (RAG), Fine-tuning, Low-Resource Languages, Medical Artificial Intelligence, UlizaLlama, Generative Language Models

Pages: 102-122