Definition
Grounding is making a language model answer from specified material rather than from what it absorbed in training. A language model writes by predicting likely text, as the guide to how language models generate text explains, so without the facts in its input it produces something with the shape of an answer. Grounding changes the input, not the model.
It comes in two forms. Manual grounding is a source pack: someone gathers the specification, the price list and the approved claims, pastes them in and instructs the model to use only those. Automatic grounding, usually called retrieval or retrieval-augmented generation, does the gathering in software for every question. It sits behind most help-centre assistants and "ask your documents" features.
In practice: indexing, retrieval and source-based prompts
The guide to how grounding and retrieval work follows the chain step by step. In outline:
- A library is chosen: the help articles, product pages and policies the tool may draw on.
- Documents are split into passages and indexed, often by meaning and sometimes also by matching words.
- For each question, the best-matching passages are found and placed in the model's input.
- The instructions set the rules: answer only from these passages, cite each point, and say so when the passages do not contain the answer.
The guide to accuracy and review workflows adds a source hierarchy for conflicts: the business's own structured data outranks its own prose, which outranks public primary sources, which outrank anything the model "knows". In customer service, the guide to AI in customer service makes approved content the assistant's limit, which turns the help centre into something that needs an owner and an update routine. The AI marketing hub lists grounding among the mechanics a marketer needs before relying on a model's output.
What grounding reduces and what it cannot guarantee
Grounding reduces invented facts and keeps answers current: change a document and the next answer can change with it, with no retraining. It also makes answers traceable, because each one can point to the passages it was given.
It cannot guarantee correctness. Failures fall into three groups. The right passage is never found, because the question uses different words or the fact was split away from its heading. The wrong passage is found, such as an old version of a policy that matches the question's wording better. Or the model strays from what it was given, combining two passages into a claim neither makes. A citation shows what the model was given, not that a sentence is supported, so answers that matter are still checked against their sources.
A travel insurer's assistant stops inventing cover limits
Pacthollow Travel Insurance piloted a website assistant that answered from the model's general knowledge with a short brand prompt. In a review of 200 test questions, it gave a specific baggage cover limit in 23 answers, and none of the figures matched any of the insurer's 4 current policy wordings. The team rebuilt the assistant to answer only from passages retrieved from the current policy documents, cite the policy and section, and hand over to an agent when no passage covered the question. On the same 200 questions, invented limits fell to none and 31 questions were handed over. Reviewers now check each answer against the passages it cites, which is faster than searching the whole library, and the 31 hand-overs showed which policy sections customers could not find.
Related terms
A Language ModelAn AI system trained on large amounts of text that writes by predicting which piece of text is likely to come next. is what grounding constrains: the writer that otherwise fills gaps with plausible text. An EmbeddingA list of numbers that represents a piece of text so that texts with similar meaning sit close together, used for search, retrieval and clustering. is the usual way passages are indexed and matched to a question by meaning. Fine-TuningTraining an existing language model further on a set of examples so that a style, format or task pattern is built into the model itself. is the alternative that changes the model itself; it shapes style and format but carries facts only as an impression, which is why facts that change belong in grounded documents. A change to the library or the instructions is best treated as a HypothesisA testable statement of the change to make, the effect expected on a named metric and the evidence-based reason for expecting it., tested against a fixed set of real questions before it goes live.
Common confusions
- A grounded tool cannot invent. It invents less; it can still misread, blend or ignore what it was given.
- A citation proves the sentence. It shows which passages were supplied, not which sentence came from which.
- Grounding is a model setting. Most of its quality comes from the documents: duplicates, stale versions and passages that make no sense on their own cause most wrong answers.
- Grounding and fine-tuning do the same job. One supplies facts at the moment of answering; the other changes how the model writes.
- Retrieval is only for large teams. A maintained folder and a prompt that says "use only these sources" is grounding too.
