Grounding
Connecting a model's output to trusted sources, retrieved evidence, tools, or real-world state.
Which gallery entrance can the assistant name?
A fictional weekend notice reads, “During roof work, visitors enter through the east door.” A visitor asks which entrance to use and whether it has step-free access.
Supported answer
The notice directs visitors to the east door during the roof work. Link the answer to that notice and identify the period it covers.
Unanswered detail
The notice says nothing about steps. Check a current access guide or ask the venue before making a claim about the route.
When retrieved passages disagree
An older floor plan may mark the west door as the usual entrance. Record the dates and purposes of both sources. A temporary notice can answer a question about this weekend while a permanent access guide may be needed for the second part of the visitor’s question.
How Grounding Is Implemented
One common method is RAG, where the system retrieves relevant passages before generation. Another is tool use, where the model queries a trusted API or database instead of guessing. Some systems also enforce citations, source snippets, or output validation to keep the answer closer to evidence.
Grounding does not eliminate error, because the retrieved material may still be incomplete, noisy, or misunderstood. But it is a major step toward more auditable and useful AI systems, especially in professional settings.
Follow the idea
RAG · Retrieval-augmented generation can bring a relevant notice into the model’s context.
Evidence connects each part of the answer to material that supports it.
Related concepts (5)
Examples in the AI guides
More chapter examples (2)
Related Yenra articles (4)
Source · Microsoft documentation on evaluating grounded answers
Return to Grounding in the A–Z glossary