Empathetic AI for Mental Health: A Comparative Study of Retrieval-Augmented Language Models in CBT Applications

The global mental health crisis has underscored the urgent need for scalable, accessible, and cost-effective support systems. This study presents MindHaven, an AI-driven chatbot designed to deliver personalized mental health support based on Cognitive Behavioral Therapy (CBT) principles. Leveraging state-of-the-art Large Language Models (LLMs), including Llama-3.2-3B-Instruct, BlenderBot, and Retrieval-Augmented Generation (RAG)-enhanced architectures, we systematically evaluate chatbot efficacy in generating empathetic, contextually relevant, and clinically informed responses. A rigorous comparative analysis was conducted across twelve transformer-based models, fine-tuned on 99,086 structured therapeutic dialogues. Model performance was assessed using quantitative NLP metrics (ROUGE, BLEU, Semantic Similarity) and qualitative human evaluations. Results indicate that Llama-3.2-3B-Instruct achieved the highest semantic similarity score of 0.8123, while DistilGPT-2 with RAG demonstrated a 15.4% improvement in response factuality over non-RAG models. Additionally, Llama-3.2-3B-bnb-4bit obtained the lowest training loss (0.6051), significantly outperforming larger models such as MedAlpaca-7B. These findings suggest that parameter-efficient architectures can rival larger models in mental health applications. Despite promising advancements, challenges remain in long-term context retention, ethical safeguards, and crisis intervention capabilities. Future work will focus on multi-lingual generalization, dynamic memory augmentation, and psychologist-in-the-loop validation to ensure clinical safety and efficacy. By bridging cutting-edge AI with evidence-based mental health interventions, this research lays a foundation for the responsible deployment of scalable, AI-assisted therapeutic solutions, with potential applications in digital psychiatry, mental wellness apps, and hybrid human-AI therapy frameworks.