Cortexa AI Glossary · Asking well
What Is Grounding?
From Cortexa Learn, by Cortexa Consulting. Last checked .
Listen · From Cortexa Learn
Tip: once it’s playing, tap any line to jump the audio there.
Hand a chatbot the document first, and its answer changes. That move has a name.
Paste it in first
You want to know whether your apartment lease lets you leave early. You could ask a chatbot and let it answer from memory. Or you could paste the lease into the chat and ask, “Using only this document, when can I leave early?” The second answer deserves more of your trust, and the reason has a name. Tying an artificial intelligence (AI) answer to a source you can check is called grounding.
Memory or a source
A language model answers from patterns it picked up in training. When it's unsure, it doesn't stop. It writes the likeliest-sounding answer anyway, in the same confident tone it uses for facts. Grounding changes where the answer comes from. Google Cloud's guide for developers describes it as anchoring a model's responses to specific information, documents, and data sources. The model still does the writing. The facts come from the source in front of it.1
Three ways in
You'll meet grounding in three forms.
- You hand it the source: you paste or upload a document and ask about that document.
- The tool looks it up: retrieval-augmented generation (RAG) searches a set of documents and passes the best passages to the model.
- The tool searches the web: some chatbots and search engines run a search, then write from the pages they found.
Why it helps
Grounding helps in two ways. It cuts down on made-up answers, because the model works from the page instead of from memory. And it makes the answer checkable. Anthropic's guidance for developers suggests asking the model to quote the exact sentences it relied on, then to take back any claim it can't find a quote for. A quote you can find in the document is a claim you can verify in seconds.123
Grounded in the wrong place
Grounding is only as good as its source. Hand the model last year's lease, and you'll get a confident answer about the wrong lease. A search can pull an out-of-date page, too. So here's a fair test for any grounded answer: does it use what the source says, and does it stay inside it? When a reply adds a detail the source never mentions, that detail came from somewhere else.
What it doesn't fix
Grounding aims at facts. It does nothing for a tone that's off, or for advice that's wrong for your situation even when every fact in it is right. So it works best alongside other checks, like a person who reads the answer before it goes anywhere important.
Ask for the lines
You can ground an answer yourself today. Give the chatbot the document, ask it to answer only from that document, and ask it to quote the lines it used. Then find one of those lines in the original. If it's there, and it says what the answer says, you're on solid ground. Which document on your desk would you want answers from this week?