Running summarizing tools on a new article

Abstract: Background: The integration of large language models (LLMs) in clinical nutrition recommender systems is limited by computational complexity, privacy concerns with patient data and a lack of mechanistic explanation. Small language models (SLMs) with fewer than 4 billion parameters are available for edge deployment, but their use in clinical reasoning in domain specific nutrition problems is not yet explored especially in India. Objective: This study evaluates and compares the clinical reasoning performance of four SLMs (Phi-3.5-mini-instruct, Phi-4-mini-instruct, Qwen2.5-3B-Instruct and Llama-3.2-3B-Instruct) in a Retrieval-Augmented Generation (RAG) framework for customized Indian clinical meal planning. Methods: The proposed framework integrates a Qdrant vector database with BAAI/bge-m3 embeddings and a four-step clinical reasoning prompt with condition identification, dietary constraint definition, food selection rationale and explainable AI (XAI) output generation. We tested Qwen2.5-3B-Instruct performance on 50 synthetic patient profiles of 10 chronic disease conditions using three evaluation metrics: Chain-of-Thought compliance (CoT), XAI keyword recall (XAI) and RAG Adherence and Fidelity (RAF). We further evaluated Qwen2.5-3B-Instruct with five prompt engineering tasks: clinical query rewriting, clinical reranking, safety filtering, forced citation formatting and few-shot exemplar injection. Results: Under the full prompt engineering configuration, Qwen2.5-3B-Instruct showed strong structural compliance (mean CoT = 0.975; 96% perfect scores) and strong RAG grounding (mean RAF = 0.960). Nevertheless, there was a clinical depth gap in the XAI outputs (mean XAI = 0.616), which indicates that although our model does well to keep the retrieved evidence and reasoning structure in line, mechanistic clinical explanation remains a major limitation. Conclusions: These results demonstrate the feasibility of using sub-4 billion-parameter SLMs for edge-deployed privacy-preserving clinical nutrition systems and provide practical optimization targets to further research as well. The XAI deficit points toward prompt and architecture-level interventions as potential ways to bring explainability in SLM deployment particularly for meal recommendation systems.