Mindmatrix: A Retrieval-Augmented Generation Framework for Multi-Class Mental Health Risk Assessment Using Large Language Models

Mental health disorders have emerged as a major global concern, affecting millions of individuals and creating a growing demand for accessible and intelligent assessment systems. Traditional mental health assessment methods often rely on manual evaluation, clinical expertise, and time-consuming diagnostic procedures. Recent advancements in Artificial Intelligence, Natural Language Processing, and Large Language Models have opened new possibilities for developing automated systems capable of analyzing textual narratives and identifying potential mental health risks. This paper presents MindMatrix, a Retrieval-Augmented Generation (RAG) framework designed for multi-class mental health risk assessment using semantic retrieval and Large Language Models. The proposed system integrates Sentence Transformer embeddings, ChromaDB vector databases, semantic similarity retrieval, and the LLaMA 3.1 language model to generate context-aware and explainable mental health assessments. User narratives are transformed into semantic embeddings and compared against a knowledge base containing mental health records. Relevant narratives are retrieved and provided as contextual evidence to the language model before generating an assessment. The framework supports multiple mental health categories including Depression, Anxiety, Stress, Bipolar Disorder, Personality Disorder, Suicidal Ideation, and Normal behavior. Experimental observations demonstrate that the integration of semantic retrieval improves contextual understanding and enhances the quality of generated assessments. The system further provides confidence scores, risk levels, explanations, and recommendations through an interactive dashboard. The proposed framework highlights the potential of Retrieval-Augmented Generation in developing explainable and reliable mental health assessment systems.