Converging Retrieval-Augmented Generation, Agentic AI, and Digital Psychiatry for Safer Clinical Decision Support
Main article
Abstract
Large language models are increasingly being tested as clinical-facing assistants, but their direct use in psychiatry remains risky because psychiatric assessment depends on careful interpretation of patient narratives, diagnostic criteria, contextual uncertainty, and escalation rules. This article develops a future-technology framework for safer clinical decision support by integrating retrieval-augmented generation, agentic orchestration, and digital psychiatry. Instead of treating the language model as an autonomous diagnostician, the proposed architecture assigns it a constrained role within an evidence-traceable workflow. The model first extracts symptom-bearing spans from patient text, an agent converts those spans into retrieval queries, a clinical knowledge layer returns criteria and screening evidence, and a final reasoning module generates a structured screening decision linked to retrieved evidence. A safety checker then evaluates unsupported claims, contradictions, self-harm cues, and escalation requirements before the output is presented for clinician review. Drawing on the benchmark structure of a public depression-detection evaluation using 100 labeled narratives and four open LLM families, the paper presents a comparative analytical assessment of direct prompting and traceable RAG-agent workflows. The analysis shows why accuracy alone is insufficient for clinical deployment and introduces evidence coverage, citation linkage, contradiction control, escalation sensitivity, and review efficiency as complementary safety metrics. The paper contributes a clinical-technical framework, a formal decision model, an evaluation matrix, and a governance roadmap for RAG-agent psychiatry systems.
