AI 101

Grounding

Also known as: grounded generation

Inference

Grounding is constraining the model to answer from the sources you give it rather than from its (possibly wrong) memory. It’s the core antidote to hallucination.

A grounded answer says “according to this document…” — and can point to where.

  • RAG is the main grounding mechanism.
  • Citations make the grounding checkable.
  • “I don’t know” is grounded honesty when the source doesn’t cover it.

The discipline: for factual answers, always ground. Ask the model to cite its source and to say "not in the provided material" when it's not there. That's how you buy reliability.