<![CDATA[Mental health challenges in Indonesia highlight the urgent need for accessible, safe, and responsible early psychological support systems. Recent advances in Large Language Models (LLMs) enable the development of mental health chatbots capable of generating empathetic and context-aware responses. This study proposes and evaluates a hybrid LLM-based mental health chatbot architecture that integrates parameter-efficient fine-tuning (QLoRA), short-term user profiling through emotion detection, and Retrieval-Augmented Generation (RAG) to improve response quality, relevance, and safety. The research methodology is designed to be modular and reproducible, encompassing Indonesian mental health dialogue preprocessing, QLoRA-based LLM fine-tuning, IndoBERT-based emotion recognition, and a FAISS-powered RAG framework using sentence embeddings. The contribution of each component is systematically assessed through a staged ablation study, while response quality is evaluated using BLEU, ROUGE-L, METEOR, and CIDEr metrics, complemented by qualitative analysis and safety stress testing. Results indicate that although BLEU scores remain relatively low—consistent with open-domain dialogue systems—higher METEOR and CIDEr scores demonstrate strong semantic alignment and informational relevance. Furthermore, the system consistently identifies user emotions, rejects high-risk requests, and produces non-diagnostic responses aligned with AI safety principles. These findings demonstrate that the proposed hybrid LLM–RAG architecture is effective as a context-aware, safe, and responsible early-stage mental health support system, without replacing professional clinical services.]]>
Copyrights © 2026
