The integration of generative artificial intelligence (AI) into poultry nutrition offers new opportunities to enhance decision-making, support feed design, and improve animal health.
Researchers at the University of Georgia in the US introduced NutriCHAT, a domain-specific AI agent designed to address the unique challenges of poultry nutrition. Built on a novel hybrid architecture combining the ReAct (Reasoning + Acting) framework with Retrieval-Augmented Generation (RAG), NutriCHAT represents a significant step toward precision agriculture.
Inside NutriCHAT’s toolkit
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 covering six production parameters of Ross and Cobb broilers).
Outperforming leading language models
NutriCHAT was evaluated against GPT-4o and four other large language models (LLMs): Grok-beta, GPT-4o mini, GPT-3.5 Turbo, and Gemini 2.0 Flash. Using 120 queries assessed for correctness, precision, and scientific depth metrics (5-point Likert scale), NutriCHAT achieved an overall weighted improvement of 83.87% compared to GPT-4o, including gains of 72.18% in correctness, 163.98% in precision, and 52.31% in scientific depth.
EvaluatorLLM corroborated these findings with 25.86% overall improvement, while SelfCheckGPT-NLI (N = 20) hallucination analysis revealed a 61.90% reduction in hallucination propensity (0.072 vs 0.189) compared to the baseline GPT-4o model.
NutriCHAT also achieved superior readability (44.9 Flesch Reading Ease score) compared to the five LLMs, demonstrating its dual benefits of accuracy and accessibility.
Architecture that drives results
Ablation studies confirmed the complementary roles of architectural components. The ReAct framework primarily drove improvements in correctness and precision, while the RAG component enhanced scientific depth by 48.31%.
Tool-specific analysis showed particularly strong performance in data-intensive queries, with precision gains of up to 187.36% in amino acid composition tasks.
NutriCHAT also proved economically viable. Operating at an average cost of USD 0.0124 per query with inference times of 6.89 seconds, it offers a cost-effective alternative to continuous expert consultation.
Its modular architecture allows straightforward integration of additional nutrition tools and potential extension to other livestock species through curated domain-specific databases.
This study establishes the feasibility of a knowledge-grounded LLM agent for precision agriculture applications, advancing data-driven poultry nutrition management by providing researchers and farmers with reliable, scientifically grounded answers from authoritative knowledge bases while mitigating the hallucination propensity, output uncertainty, and data trust challenges that limit general-purpose LLMs in specialized agricultural domains.