A RAG chatbot built for a network of 25,000+ implant dentistry specialists achieved 98% accuracy and near-zero hallucinations using GPT-5 mini — a deliberately small model. The key insight: in RAG systems, accuracy is determined by data preparation and retrieval quality, not model size. The engineering effort went into rigorous source cleaning (handling clinical abbreviations, measurement formats, hyphenated breaks), empirically-tested chunk sizing (1,000–1,250 tokens to keep complete answers intact), strict prompt constraints (no fabrication, brand-neutral, scope-limited), and a golden dataset evaluation pipeline using Promptfoo and Langfuse. User testing with 24 specialists and a 31-member survey returned an NPS of 81, with 23 of 24 specialists clearly preferring it over general chatbots due to traceable, vetted citations.
