WTF is Grounding?
Tying a model's answer to verifiable sources so it states facts it can back up instead of guessing.
Also called: grounded generation
TL;DR
An ungrounded model answers from memory and can confidently make things up. A grounded one is given real sources (your documents, search results, a database) and answers from them, ideally with citations you can check. RAG is the most common way to do it.
How it works
Google's definition: connecting a model's output to verifiable sources of information to improve accuracy and reduce hallucinations. Implementations include retrieval over private data, live web search, and structured lookups via tool calls. Good grounding also means instructing the model to say when the sources don't contain the answer, and returning citations so users and evals can verify claims.
Origin
- Coined / introduced by
- Industry term (popularised by cloud AI platforms)
- When
- 2023
- Story
- Became standard vocabulary as vendors shipped search- and data-grounded generation features.