Ground responses in your sources
- prompting
- verification
Grounding means the answer comes from material you supply — not from what the model happens to remember. Instead of trusting the engine’s training memory, you paste the source and constrain the answer to it. The payoff is enormous: every claim becomes checkable against a text you can hold side by side.
The grounding pattern
- Paste the source — the policy, the paper, the contract. Not a link, not a description.
- Constrain to it — “Answer using only the text below.”
- Demand traceability — “For each statement, quote the sentence it comes from.”
- Define the fallback — “If the text does not cover something, say ‘not covered’.” Without this line, gaps get filled with confident invention.
- Verify by comparison — you check quotes against the source. That check is cheap, which is the whole point.
A bad example
Ask: “What does our refund policy say about change-of-mind returns?”
The model has never seen your policy. You get the average refund policy of its training data — fluent, structured and wrong in exactly the details that matter.
A better example
Ask: “Here is our refund policy: [paste]. Answer this question using only the policy text. Quote the sentence behind each statement. If the policy does not cover it, answer ‘not covered’.”
Now the answer is a set of pointers into your document. Wrong answers are visible: either a quote does not exist, or a claim has no quote behind it.
Why it works
The mechanism from what an LLM actually does applies here: plausibility is generated, truth is not. Grounding does not make the model truthful — it moves errors into a place where you can catch them cheaply. Support turns into a verifiable property: does this sentence appear in the source?
A warning that keeps it honest: grounding does not eliminate mistakes. Quotes can be misquoted, cherry-picked or subtly altered. Read quotes against the source rather than trusting the quotation marks.
Grounding scales as a habit, not just a trick: paste source, constrain to it, demand quotes, define the fallback, compare. Done for every high-stakes answer, it stops feeling like extra caution and starts feeling like speed — verification becomes reading a quote instead of re-reading a document. The one-line version for small questions: “answer only from the text I gave you.”
Practice
Spot the unsupported policy claims
You pasted your refund policy into a chat and asked whether change-of-mind returns are allowed. The AI answered — compare every sentence of the answer against the policy excerpt you supplied.
Select every part of the AI answer that is NOT supported by the pasted policy text.
The AI answer
POLICY (excerpt you pasted): “Refunds are available within 30 days of purchase for unused items in original packaging. Shipping costs are non-refundable. Change-of-mind returns follow the same 30-day window.” AI ANSWER: Yes — you can return unused items within 30 days in original packaging. Change-of-mind returns follow the same window. Shipping costs are non-refundable, unless the item arrived damaged, in which case shipping is refunded. Refunds are processed within 5 business days.
Hint
Compare each sentence to the excerpt: does its meaning appear there?
Specifics absent from the source — exceptions, timelines, numbers — are the suspects.
Why this is the answer
Grounded answers can be checked line by line: supported claims trace back to the source, invented extras read just as fluently as the real ones. The skill is comparing each statement to the pasted text — not trusting the tone of the answer.
Transfer
- Contracts and policies: quote-first answers; anything unquoted is flagged for a human.
- Research notes: same pattern against each paper — claim, quote, page.
- Support macros: ground replies in the official policy text, never in vibes.
Next
Next: Structured output (JSON, simply) — ask for answers shaped for your tools instead of another wall of prose. The grounding glossary entry stays the one-line version of this lesson.