Modern AI systems are increasingly used to support tasks that rely on scientific evidence, precise terminology, and contextual interpretation. Large language models (LLMs) are employed in such implementations, providing flexible and robust natural language interfaces. However, LLMs alone often struggle to meet system requirements. Due to their statistical nature, the answers produced can demonstrate hallucinations, lack awareness of their provenance, and a tendency to drift beyond the boundaries of scientific knowledge. We present a semantic architecture for building evidence-aware AI systems, where ontologies and knowledge graphs (KGs) define the knowledge layer that governs how LLMs interpret user questions and compose answers.

Using mental and behavioral health as a motivating example, we describe how instrument semantics (including their computable structure and associated mental health constructs), study data supporting measurement instruments and treatments, and clinical behavior coding systems can be integrated into a unified KG. A conversational agent built on this graph uses chat intents, SPARQL queries, and structured fact grounding to ensure that all authoritative content originates from verified sources, with the LLM serving as a natural language interface. We demonstrate how ontologies can function as an interface for machine learning, enabling safer, context-sensitive reasoning. We conclude by discussing how this pattern may be adapted to other domains.