Grounding
Also known as: Content grounding
In short
Connecting an AI response to identifiable information sources so its factual claims can be checked against evidence.
Grounding connects an AI response to information that can be checked. For a merchant, that might mean answering product questions using approved specifications or current inventory records.
Retrieval-augmented generation is one way to supply that information. Grounding is the broader goal of anchoring claims in evidence, rather than relying entirely on information represented in a model’s parameters.
A practical example
Imagine a shopper asking whether a dining table is available for delivery this week. The assistant needs the correct product variant, current stock, delivery destination, and the applicable shipping information.
A product description that says “fast delivery” is insufficient evidence for a specific arrival date. In this illustrative case, a grounded answer reports what the available records establish and leaves the delivery promise unresolved when essential information is missing.
Evidence still needs checking
Grounding reduces the risk of unsupported output. It does not eliminate errors, often called hallucinations. A system can retrieve the wrong product, use an old source, or add a claim the source never made.
Check the connection between each important claim and its evidence. Prices, stock, warranties, and compatibility deserve particular attention because small errors change a customer’s decision. A confident tone or a visible citation cannot replace that check.