Food additive lens: an on-device AI application for real-time science-based consumer education on food additives using retrieval-augmented generation

Abstract

Consumer concerns about food additives have intensified amid widespread misinformation, with the 2024 IFIC survey revealing that 35% of consumers actively avoid artificial ingredients despite authoritative safety data existing in FDA and USDA databases. This work investigates whether on-device artificial intelligence can effectively translate complex regulatory information into accessible consumer education while maintaining scientific accuracy and privacy. This paper presents Food Additive Lens (FAL), an iOS application implementing a three-agent architecture: (1) a food category classifier achieving 87.2% top-3 accuracy across 257 categories, (2) a hybrid additive identifier combining database lookup with AI extraction (F1-score: 0.757), and (3) an explanation generator producing contextualized, consumer-friendly descriptions. The system deploys Meta's Llama 3.2 3B model quantized to 1.8 GB through 4-bit compression, achieving a generation speed of 13–30 tokens/second while operating entirely offline. Integration of FDA's Substances Added to Food Inventory (3971 substances) and USDA's Global Branded Food Products Database enables comprehensive coverage with direct links to the Code of Federal Regulations for professional users. The Retrieval-Augmented Generation workflow grounds AI responses in authoritative sources, reducing hallucination while maintaining accessibility. Performance evaluation on iPhone 14 and MacBook Air M1 demonstrated stable memory usage (peak: 2.36 GB) with complete offline functionality, ensuring user privacy. The application transforms complex ingredient lists into accessible information through camera-based OCR scanning, progressive disclosure interfaces, and context-aware explanations tailored to specific food products. This work demonstrates the feasibility of deploying sophisticated AI for science communication on consumer devices, offering a scalable model for combating food-related misinformation while preserving privacy and accessibility.