Skip to main content
Glossary

Grounding

Definition

Grounding means tying a language model’s output to identifiable source material, so that statements can be traced back to the evidence they came from rather than resting on the model alone.

Related
Citations, attribution, RAG
The question to ask
Grounded to what granularity?
Useful level
The verbatim sentence

Grounding is a spectrum, and most of it is useless​

The claim is nearly meaningless without a granularity, because all of these are technically grounding:

What the answer gives youCost of checking one claim
  1. “Based on your documents”Impossible to check
  2. A list of file namesOpen each file, read until you find it
  3. The document and a page numberFind the page, scan it
  4. The verbatim sentence, one click from its pageRead one sentence

Only the last row is cheaper than re-reading the source yourself — which is the only reason anyone actually checks.

The same claim, grounded four ways. The bar is what checking it costs you.

Only the last row survives contact with real work. Anything above it costs more than re-reading the source yourself — so nobody does it, so the citation stops functioning as a check and becomes decoration. A citation that is never followed is worse than no citation, because it buys unearned confidence.

The second thing granularity hides: a real quote can be attached to a conclusion it does not support. The source exists, the page number is right, and the inference is wrong. Grounding at document level cannot detect that. Grounding at sentence level makes it obvious in a click.

Two questions worth asking any tool​

  1. Can I get from a claim to the exact words in under five seconds? If not, the grounding is for the vendor’s marketing, not for your verification.
  2. Is anything checked before I see it, or is the citation just attached? Attaching a source to generated text is a formatting step. Verifying that the quoted text exists in the source is a different and much rarer thing.

How MindMax handles it​

Every claim carries the verbatim quote it rests on. As the answer is finished, each quote is matched against its source and aligned to the exact wording there; one that has been paraphrased, smoothed or invented cannot be found, and is flagged in the answer rather than passed off as checked. The mechanism is described in checked quotes.

Clicking a quote opens the document at that page and highlights the sentence.

One citation, opened: the document, the page, and the exact sentence — highlighted in the PDF on the left.

Common questions

What does grounding mean for an AI answer?

That its statements can be traced to identifiable source material. How useful that is depends on the granularity: a filename tells you where to start looking, while the verbatim sentence lets you check the claim in seconds.

Are citations the same as grounding?

A citation is one form of grounding. Whether it is a useful one depends on what it points at and whether anything checked it — a citation attached to generated text without verifying the quote is formatting, not evidence.

How do I check a grounded answer quickly?

Follow the claims that the rest of your work depends on, not all of them. If each points at an exact sentence one click away, checking the three that matter takes a minute.