noun · AI technique
Grounding
also called source grounding
Making the AI show its sources. Instead of answering from memory, it answers from material you gave it and can point back to.
Anchoring a model's answer in supplied sources, such as retrieved documents or tool results, so claims can be traced and checked.
The harness supplies documents, search results or tool output and instructs the model to answer from them, often with citations. The model still generates the text, so the link to the source has to be checked, not assumed.
Real-life example
You ask a help bot about a refund rule. Without sources, it gives a smooth answer from memory. With grounding, it first pulls the policy page, answers from it, and quotes the sentence it used. You can check the quote against the page in seconds.
Where the analogy breaks
-
Grounded does not mean true. The source itself can be wrong or out of date, and the answer will repeat it faithfully.
-
The model can still add things the source does not say. A citation shows where to look, not that the claim matches.
-
Grounding depends on finding the right source. If retrieval brings back the wrong page, the answer is confidently grounded in the wrong thing.