Grounding
Tying a model's answer to retrieved sources, so that its claims come from text the system actually found.
What it means
Grounding means the answer is built from material the system retrieved, not only from what the model learned during training. A grounded answer can point to the pages it used, and a well-built system checks that those pages support each claim. Grounding reduces some kinds of error, but it does not remove them. A model can still misread a source, merge two passages, or stretch a sentence beyond what it says.
When you evaluate a tool, ask how grounding works. Does the tool quote the source? Does it check quotes against the page? Does it return a 'not found' result when the search fails? Those answers say more than a general claim that the model is accurate. Related terms are retrieval, citation and hallucination.
Continue this reading path
Cite this entry
https://bestaisearchengine.com/glossary/grounding/