A grounding instruction constrains the model to use only the context explicitly provided — retrieved documents, user-supplied data, or tool outputs — rather than its parametric knowledge. This technique is fundamental to RAG pipelines and enterprise deployments where factual accuracy matters more than creativity, and is often combined with citation requirements so users can verify claims.