Grounding
Last updated:
What grounding means
In AI, to ground an answer is to base it on evidence the system can point to. That evidence can be retrieved documents, a database record, the result of an API call or facts the owner has configured. The opposite is an ungrounded answer, which comes only from what the model absorbed during training and may be outdated, generic or invented.
Grounding is closely related to retrieval-augmented generation. Retrieval supplies the evidence; grounding is the rule that the answer must stay within it.
How grounding works
In a chatbot, grounding usually combines four elements:
Grounding is a matter of degree. A chatbot may still use general knowledge to explain a concept, as long as business-specific facts come from approved sources.
- Evidence: the system retrieves relevant passages or calls a tool and places the results in the prompt.
- Rules: instructions tell the model to use that evidence for factual claims, not to invent prices, links or policies, and to say when evidence is missing.
- Source ranking: when sources disagree, the system defines which one wins, for example an approved Q&A answer over an older web page.
- Checks: some systems show citations or compare the answer with the sources afterwards.
Example
A dental clinic’s chatbot is asked: “Do you accept my insurance, and how much is a cleaning?” The retrieved page lists accepted insurers and the cleaning price, and a grounded answer repeats both exactly as written. If the visitor names an insurer that is not on the list, the grounded answer says the page does not mention it and suggests calling the clinic, instead of guessing yes or no.
Why grounding matters for business chatbots
Customers treat a chatbot’s answers as the company’s word. Grounding makes those answers specific to your business, consistent with your published information and easier to audit, because every fact should trace back to a source.
It also sets a clear limit. If your content is wrong or incomplete, a grounded chatbot repeats the error or admits the gap. Good grounding therefore depends on good content.
How intoCHAT grounds answers
intoCHAT applies built-in grounding rules to every agent, and they take precedence over conflicting custom instructions. Business facts such as products, prices, policies and contact details must come from your configuration, your retrieved content or a successful action result. Retrieved excerpts are ranked: Q&A answers first, then starred sources, then text snippets, documents and web pages, with newer content winning between equals. When the evidence is missing or conflicting, the agent says what it cannot confirm instead of inventing an answer.