Emerging Trends in Retrieval-Augmented Large Language Models for Mental Health Assessment

Abstract

Background: The convergence of Retrieval-Augmented Generation (RAG) and Large Language Model (LLM) technologies is reshaping the landscape of computational psychiatry, offering new ways to support depression screening, symptom triage and decision support in mental health practice. Objective: This article surveys emerging trends in retrieval-augmented LLM systems for mental health assessment, synthesises a representative two-stage agent-based instantiation of these ideas, and reports a comparative empirical analysis on four open-weight LLM families. Methods: We organise the literature into a three-branch taxonomy (single-pass RAG, agent-orchestrated RAG and hybrid retrieval RAG) and instantiate the agent-orchestrated branch using a two-stage symptom-to-evidence pipeline grounded in clinical practice guidelines. Four LLMs (Gemma-3, Qwen-3, DeepSeek-R1, Llama-3.1) at the 4–8B parameter scale are evaluated on a public depression detection dataset under baseline (direct querying) and augmented (RAG-Agent) conditions. Results: The augmented condition delivers improvements in accuracy of up to 17 percentage points (Llama-3.1: 57% to 74%) and in precision of up to 17 percentage points (Gemma-3: 76.81% to 94.12%), accompanied by modest, model-dependent reductions in recall. The framework consistently produces citation-supported outputs that align with explainable artificial intelligence requirements. Conclusion: Retrieval-augmented LLMs combined with structured agent reasoning constitute a viable, scalable, and clinically interpretable pathway for AI-assisted mental health screening, with practical implications for translational research and biomedical engineering.