Personalized nutrition coaching applications aim to provide context-aware dietary recommendations aligned with an individual's health profile, medical constraints, and consump- tion habits. Traditionally, most nutrition analysis engines have relied on modular, rule-based, or ingredient-level classification architectures that process each component of the food label in isolation. With the rise of large language models (LLMs), it has become possible to construct "holistic" architectures capable of processing the entire contextual dataset—ingredients, nutrition table, and user profile—in a single inference call. This paper presents a comprehensive comparative study between two distinct design philosophies: (1) an atomistic Cache-and-Curate architecture based on decomposed multi-step classification, rule chaining, and database lookups; and (2) a holistic Single-Call Generative Architecture inspired by modern LLM capabilities. Using a gold-standard dataset annotated by registered dietitians, we evaluate accuracy, personalization depth, semantic coher- ence, latency, and operational cost. Results demonstrate that the holistic architecture significantly outperforms the atomistic pipeline across all metrics, achieving higher expert alignment, faster inference, richer user-specific reasoning, and substantially lower engineering overhead. The study provides one of the first empirical analyses of architectural paradigms for generative-AI- powered health applications and offers insights into the future direction of real-time contextual reasoning systems.
Singh et al. (2026) studied this question.