Grounded answers

In one paragraph

A grounded answer is one an AI assistant produces only after finding the right passage in your own documents, rather than writing from memory. The assistant retrieves the relevant material first, then answers using only what it found, and prints the source alongside the answer: which document, which section, and which version. This approach is sometimes called retrieval-augmented generation (RAG). When nothing relevant turns up, a grounded assistant is built to say so, rather than filling the gap with something plausible sounding.

Why it matters

A language model on its own can produce a fluent, confident answer that is simply wrong, because it is writing from patterns learned during training rather than from the document in front of it. That is a serious problem for any use where the answer needs to be checked against a real source, which covers most work inside an organisation. Grounding an assistant in your own documents, and printing the source with every answer, means a reader can check the claim in seconds rather than trusting it on faith.

How it works

• The assistant first searches your documents for the passages most relevant to the question, before it writes anything.
• It then answers using only what was retrieved, rather than drawing on general knowledge learned during training.
• Every answer carries a citation: the document it came from, the section within it, and the version, so it can be checked against the source directly.
• When the retrieval step finds nothing relevant, a grounded assistant is designed to say the answer is not in the documents, instead of producing something that sounds right but is not supported.
• A widely cited peer-reviewed study set out the common ways these systems can fail in practice; it is worth reading in full rather than repeating here, and the link is below.

What it looks like in practice

Take an assistant answering questions about a set of internal policies. Asked something the policy documents do not cover, a grounded assistant reports that it could not find an answer in the documents, rather than guessing at what a sensible policy might say. Asked something the documents do cover, it answers with the relevant passage quoted and the document and section named, so anyone reading the answer can go straight to the source and check it themselves.

How this connects to our work

Grounded answers are the basis of the company knowledge AI we build, and they depend on the governed layer underneath it, which keeps the documents an assistant reads current, permissioned, and trustworthy enough to answer from.