All terms

Grounding

Connecting model outputs to verifiable sources — documents, tools, or databases — so answers stay tethered to evidence.

Large Language Models1 min read

Definition

Grounding means the model's claims are supported by retrieved context, tool results, or structured data — often with citations users can check.

RAG is a common grounding pattern; browsing tools and SQL agents are others.

In simple terms

Speaking with footnotes — every strong claim points to a source on the desk, not just memory.

Where you see it

  • Doc chatbots that cite page numbers.
  • Customer support answers limited to policy PDFs.

How it works

  1. 1.Retrieve evidence

    Pull relevant chunks or tool outputs.

  2. 2.Condition generation

    Prompt the model to use only that evidence.

  3. 3.Attribute

    Show citations or quotes alongside the answer.

Why it matters

  • Grounding is the practical antidote to unconstrained hallucination in enterprise AI.

Often confused

  • If context is provided, the model cannot hallucinate.

    Models may still ignore or contradict context — evaluate faithfulness explicitly.