Skip to content
BENIBIX Learn

Grounding (AI)

Also known as Grounded Generation, Source Grounding

Tying a model's answer to trusted sources, so each claim can be checked.

Think of it like

Think of it like you must point to the page in the book that says so.

Example

Jack's legal tool answers with a quote and a page number for each point. When the contract says nothing about a topic, it says it found nothing.

Why it matters

It turns 'the AI said so' into 'the document says so, here', which a person can check.

Definition

Grounding is tying a model's answer to a trusted source, so each claim can be traced back to it. The source can be documents found by search, a database or a web page. A grounded answer sticks to what the source says and shows where each fact came from. If the source does not cover the question, a well-grounded system says so instead of guessing. It cuts made-up answers, but it does not end them.
Two sentences link to passages in the source. The third has no source, so it is marked to check.

Common mistake

Trusting the citation without opening it. Models sometimes cite a real source for a claim it does not make.

Key takeaways

  1. 1
    Tied to a source. Claims can be traced.
  2. 2
    Says when it does not know. No guessing.
  3. 3
    Still check. Citations can be wrong.

Compare

Learn it properly