Grounding

A model's internal knowledge is frozen at training time, lossy and unverifiable. Grounding swaps memory for evidence: retrieve the relevant passages (RAG) or search results, put them in the context window, and require the answer to cite them. The answer becomes checkable — click the citation, read the passage, judge for yourself.

The distinction from plain RAG is the requirement, not the plumbing: grounded systems refuse or flag when sources do not support an answer, and their citations actually point at the supporting text. This is the core trust mechanism behind Perplexity, NotebookLM and enterprise assistants, and the reason “chat with your docs” products live or die on their retrieval quality.