NutriCHAT: A Reasoning-Driven large language model agent with Expert-Designed tools for Knowledge-Grounded poultry nutrition Assistance

• The first large language model agent was developed in the poultry nutrition domain. • NutriCHAT achieved 83.87% improvement over GPT-4o in expert evaluation. • Four specialized tools integrated 17,017 nutrition parameters from trusted sources. • Operating cost of $0.0124 per query with 61.90% lower hallucination than GPT-4o. The integration of generative artificial intelligence (AI) in poultry nutrition can enhance decision-making, support feed design and improve animal health. This study introduces NutriCHAT, a domain-specific AI agent for poultry nutrition, built on a novel hybrid architecture combining ReAct (Reasoning + Acting) framework with Retrieval-Augmented Generation (RAG). NutriCHAT incorporates four expert-designed tools: a Feed Ingredient Bank (integrating 12,100 nutritional parameters from 100 feedstuffs and 20 amino acids, including composition, digestibility, and energy values), a Definitions Tool (600 poultry definitions), a Nutrient Requirements Tool (855 parameters representing Ross and Cobb broiler nutrition requirements across phases, genetic lines, and weight targets), and a Performance Management Tool (3462 parameters of six production parameters of Ross and Cobb broilers). NutriCHAT was evaluated against GPT-4o operating in standalone mode, and four additional LLMs (Grok-beta, GPT-4o mini, GPT-3.5 Turbo, and Gemini 2.0 Flash) using 120 queries assessed on correctness, precision, and scientific depth metrics (5-point Likert scale). Expert evaluation showed 83.87% overall improvement compared to GPT-4o. SelfCheckGPT-NLI analysis demonstrated that NutriCHAT-200 achieved a 61.90% reduction in hallucination score compared to GPT-4o. Ablation studies confirmed ReAct’s role in correctness and precision, and including RAG component in NutriCHAT contributed to improvement in scientific depth. Cost analysis showed an average query cost of $0.0124 for NutriCHAT, with average inference time of 6.89 s per query. This work demonstrates feasibility for data-driven poultry nutrition management by retrieving data from trusted knowledge bases, advancing precision agriculture via computational tools for sustainable production.