Grounding

Grounding means anchoring a model's answer to specific, provided sources — documents, search results, a database — instead of letting it answer from memory alone.

The difference is a journalist versus a man at a tea stall. Both will happily answer "what did the budget change for farmers?" The tea-stall answer draws on memory and confidence, and may be years stale. The journalist opens the actual budget PDF, answers from it, and cites the page. Same fluency; different relationship to sources. A grounded model is the journalist: supplied with reference material and instructed to answer only from it — and ideally to cite which passage supports each claim.

The standard delivery mechanism is RAG: retrieve relevant passages, place them in the prompt, instruct the model to stay within them and say "not in the sources" when the sources fall short. Grounding is the goal; RAG is the plumbing. Web-search modes in Gemini, ChatGPT and Perplexity are grounding at internet scale, and the citations they show exist so a human can verify.

What it buys: fresh facts without retraining, answers from private data, and verifiability — the citation turns "trust me" into "check me". What it does not buy: certainty. A model can still contradict its sources or smuggle in remembered "facts", a failure called faithfulness error — reduced but not eliminated by good retrieval, reranking, and instructions to decline when sources are silent. Grounded-ness itself is measurable in evals: does every claim trace to a supplied passage?

Where to go next#

• Full lesson: What is RAG?
• Related terms: rag, hallucination, chunking, reranking