What is Grounding?
Also called Grounded Generation, Source Grounding.
Grounding is the practice of tying a model's output to verifiable source material rather than letting it answer from its parameters alone. A grounded system retrieves evidence, instructs the model to answer only from it, and expects the answer to state when the evidence is insufficient. The goal is answers that can be checked against a source.
An ungrounded answer is produced from statistical patterns learned in training, which makes it fluent whether or not it is correct. Grounding narrows the task from recall to reading comprehension by supplying the relevant text and asking the model to answer from it. This measurably reduces fabricated content, though it does not eliminate it.
Grounding is a spectrum rather than a switch. Weak grounding supplies context and hopes it is used. Stronger forms require sentence level citations, verify that each claim is supported by a retrieved passage, and abstain when support is missing. Verification can use a second model pass, string overlap checks, or structured claim extraction.
The main failure mode is confident synthesis beyond the evidence. A model given three passages may combine them into a plausible fourth statement that none of them supports, and readers cannot tell the difference because the prose reads the same. Requiring citations at the claim level makes such extrapolation visible during review.
Grounding also depends on the corpus being right. A well grounded answer drawn from an outdated policy is still wrong, and the citation makes it more persuasive rather than less. For this reason grounding is inseparable from knowledge base maintenance and from surfacing document dates alongside the answer itself.
Key points
- Answers come from supplied evidence, not memory alone.
- Strong grounding requires citations and abstention when unsupported.
- Reduces fabrication but does not eliminate it.
- Models can synthesize claims beyond the passages given.
- A grounded answer from a stale source is still wrong.
In practice
Asked whether a plan includes phone support, an assistant retrieves the plan comparison page, finds no mention of phone support, and answers that the documentation does not state it, offering the page as its source. An ungrounded system would likely guess yes or no in confident prose, giving the reader no way to check.