Short answer
Grounding means tying an AI answer to specific documents you provide, instead of letting the model answer from its general training. Common techniques are giving it only your documents, asking for word for word quotes, and requiring a source for each claim.
What grounding looks like
Anthropic's guide on reducing hallucinations describes grounding responses in direct quotes, making answers auditable with citations, and telling the model to use only the provided documents and not its general knowledge. If it cannot find a supporting quote, it should retract the claim.
Grounding in a team bot
The documents are answers a person approved. Each reply shows which answer it came from and when it was confirmed. If no approved answer matches, the bot says it does not know instead of filling the gap.
Limits
The same guide says these techniques significantly reduce hallucinations but do not eliminate them, so critical information still needs a person to check it. A grounded bot is only as good as the documents behind it.
Related
Want early access and new checklists by email?
Leave your email and we write to you when your spot is ready. Or sign in now and the first 100 founders start right away.
Common questions
They are related. Retrieval-augmented generation is one way to build a system that looks up documents; grounding is the aim of tying answers to those documents.
It reduces it a lot but does not remove it. Test the bot with questions that have no answer in your documents.