Connecting a model’s response to supplied or retrieved information intended to support the answer.
Grounding connects a model’s response to supplied or retrieved information. Instead of relying only on what the model learned during training, an application gives it relevant context to use when answering. That context might include documentation, records or selected web content.
The source of grounding and the truth of the source are separate questions. A retrieved document can be inaccurate, incomplete or out of date. Grounding does not automatically verify its claims, resolve conflicting versions or establish that the user is permitted to see every item in the collection.
Grounding describes how an answer is supported by context; RAG is one way to retrieve context for that purpose. To evaluate the result, inspect both whether the answer follows the provided material and whether that material is suitable for the question. Those checks address different failure modes.
Takeaway: Maintain accurate source information and evaluate how it is used. Calling a response grounded should not replace checking important claims against the right evidence.
Example
A business assistant receives a current returns policy as context. It may give a well-supported explanation of the stated return window. If someone supplied an expired policy instead, the answer could follow that document closely while still giving customers the wrong information.
What to check
Keep source versions identifiable and define what should happen when evidence is missing. Review consequential answers rather than treating citations or confident wording as independent confirmation of accuracy.
For more detail, see Retrieval-augmented generation in Azure AI Search.