RETRIEVAL-AUGMENTED GENERATION WITH AGENTIC ORCHESTRATION AND SEMANTIC VECTOR SEARCH FOR EVIDENCE-BASED TYPE 2 DIABETES MANAGEMENT

Abstract

Large language models can generate fluent clinical text, but standalone models may rely on outdated training data, omit evidence, or hallucinate recommendations. This paper proposes and evaluates an agentic retrieval-augmented generation (RAG) framework for evidence-grounded decision support in Type 2 Diabetes Mellitus (T2DM), addressing these limitations through dynamic retrieval from authoritative clinical guideline corpora at inference time. The system implements a structured knowledge ingestion pipeline in which guideline documents are extracted and normalized via PyMuPDF, segmented into overlapping text chunks (900-character window, 120-character stride), encoded as dense semantic vectors using the all-MiniLM-L6-v2 sentence transformer, and indexed within a persistent Chroma vector store. At inference, a coordinated ensemble of Groq-accelerated LLM agents executes discrete, auditable reasoning subtasks encompassing structured patient data extraction, diagnostic confirmation against retrieved evidence, and treatment recommendation synthesis in accordance with current clinical standards. The architecture is validated through a representative clinical vignette and implemented as a reproducible Python pipeline compatible with Google Colab environments. Experimental results demonstrate that modular agent decomposition enhances output traceability, reduces unsupported inference, and supports alignment with evidence-based clinical guidelines. The proposed framework contributes a scalable, auditable reference architecture for guideline-grounded clinical decision support, with discussion of safety constraints, evaluation metrics, and limitations relevant to prospective clinical deployment. This work is intended exclusively for research and clinical decision-support exploration and is neither designed nor validated for autonomous diagnosis or therapeutic intervention