SOPHIALLM: DESENVOLVIMENTO E VALIDAÇÃO DE UM MODELO DE LINGUAGEM ESPECIALIZADO PARA RACIOCÍNIO CLÍNICO, DIAGNÓSTICO E EDUCAÇÃO EM MEDICINA VETERINÁRIA DE PEQUENOS ANIMAIS

Abstract

SophiaLLM is a language model specialized in small animal veterinary medicine (dogs and cats), developed entirely in Brazilian Portuguese to support clinical reasoning, diagnosis, continuing veterinary education and scientific knowledge retrieval. Unlike generalist models, SophiaLLM was designed exclusively for the veterinary domain, seeking greater precision and reduced hallucination incidence. The system combines a language model with approximately 109.5 million parameters with a Retrieval-Augmented Generation (RAG) architecture, allowing scientific knowledge to be stored and updated separately from model weights. Its knowledge base comprises approximately 500 million tokens (about 6.94 GB) of veterinary literature, tens of thousands of clinical cases, a knowledge graph containing 454 diseases, 680 entities and 9,546 clinical relations, plus a clinical reasoning engine capable of generating diagnostic hypotheses, differential diagnoses and complementary exam recommendations. SophiaLLM represents a pioneering initiative for the creation of specialized veterinary artificial intelligence, demonstrating that smaller models combined with structured knowledge bases and information retrieval mechanisms can deliver relevant performance for clinical, educational and scientific applications in small animal veterinary medicine.