ElderDocAI: Evidence-Grounded Retrieval-Augmented Generation with Reliability Gating and Care-State Adaptation for Older-Adult Care

Large language models can support caregiver-facing assistance, but ungrounded responses pose risks in older-adult care. We present ElderDocAI, a framework combining hybrid retrieval, reliability-gated generation, and rule-based care-state adaptation. Evaluation on GeriLit-Gold v1.1 (121 questions from 45 PubMed Central articles) yielded mean faithfulness of 0.754, answer relevance of 0.672, and evidence support of 0.853. The full pipeline outperformed vanilla retrieval-augmented generation (RAG) across all three metrics, while reliability gating alone showed no significant improvement. Gold evidence was retrieved for 27.3% of questions (Recall@3 = 0.273; MRR = 0.238). Among five alternative generator models, Phi-3.5 achieved the highest faithfulness (0.900–0.923) but failed on one question in both runs, and Mistral 7B achieved the highest answer relevance (0.711) and evidence support (0.817); refusal rates ranged from none to 47 of 121 questions. Care-state adaptation did not produce responses judged appropriate in the controlled evaluation. These findings highlight generator choice and retrieval coverage as important factors in evidence-grounded older-adult care assistance. Clinical effectiveness and longitudinal behavior remain unevaluated.