Fundamentals

Grounding

Grounding means anchoring the model's outputs to specific source documents or data. Instead of relying on training data, the model generates responses based on provided context. This is the core principle behind retrieval-augmented generation (RAG). Grounding dramatically reduces hallucination because the model can cite specific passages. It also enables the model to work with real-time data, proprietary information, and knowledge outside its training set. Effective grounding requires careful document retrieval and clear instructions for referencing sources.