Ask a general chatbot about your document and it will happily blend two things: what your document actually says, and what the model already "knows" from training. Most of the time you can't tell which is which — and for a contract, a policy, or a research finding, an answer that quietly draws on outside knowledge is worse than useless. If you want to make AI answer only from your own documents, you need a tool built to do that, and a way to check it's holding the line. Here's how.
Why general chatbots don't stay in the document
A model like ChatGPT is "open-book plus closed-book" by default: it reads what you give it, but it also answers from everything in its training data. So when your document is thin on a detail, the model fills the gap from general knowledge — confidently, and without telling you. That's how you get an answer that sounds right, matches the document's topic, and is actually made up. It's the same mechanism behind AI hallucination in document summaries.
What you want instead is a grounded answer: the tool retrieves the relevant passages from your document, answers only from those, and refuses to invent the rest.
How to make AI answer only from your documents
1. Use a tool that retrieves from your document first
The technique is retrieval-augmented generation: the tool finds the passages in your document relevant to your question, and the model answers from those, not from memory. A general chat box doesn't do this; a document reader built around it does.
2. Ask against the document
Put your question to the document you have open. The answer should be assembled from that document's text.
3. Require a citation on every claim
This is your enforcement mechanism. If every claim links to a passage in your document, there's nowhere for outside knowledge to hide. A claim with no citation is a claim the tool couldn't ground — and that's exactly what you want to catch.
4. Test the boundary
Ask something the document doesn't cover. A properly grounded tool should tell you it can't find the answer, not answer from general knowledge. If it confidently answers a question your document never addresses, it isn't staying in the document.
How Sidenote enforces this
Sidenote is built around exactly this discipline. It reads the document you have open, retrieves the relevant passages, and answers from them — and then it does something most tools don't: it validates every claim against the passages it actually retrieved, and drops any claim it can't match to a real passage before you ever see it. That's source-grounding in practice.
The effect is that you get one of two things: a cited answer you can click to verify against the exact sentence, or an honest "the document doesn't say." What you don't get is a confident answer quietly assembled from the model's general knowledge. For anything you'll act on — a clause, a policy, a number — that boundary is the whole point.
What "answer only from my documents" should look like
| Behaviour | General chatbot | Grounded reader (Sidenote) |
|---|---|---|
| Source of the answer | Your text + training data | Passages retrieved from your document |
| Claim with no support | Answered anyway | Dropped before you see it |
| Question the doc doesn't cover | Confident guess | "Not found in the document" |
| Verify a claim | No link to source | Click to the exact highlighted passage |
Frequently asked questions
Can I stop ChatGPT from using its own knowledge and only use my file?
Not reliably with prompting alone — a general chatbot still has its training data available and will fall back on it when your file is thin. To genuinely scope answers to your document, use a reader built on retrieval and grounding, like Sidenote, which answers from retrieved passages and drops claims it can't match to your text.
What's the difference between citing sources and answering only from sources?
Citing sources shows where an answer came from; answering only from sources refuses to use anything else. You can have one without the other — a tool might cite a passage while still blending in outside knowledge elsewhere. The strongest setup does both, which is why getting AI to cite its sources and grounding answers to your document go together.
How do I know the AI didn't make part of the answer up?
Check that every claim is cited and that the tool says "not found" when you ask something outside the document. Sidenote drops any claim it can't ground in a retrieved passage, so an uncited, invented answer doesn't reach you in the first place — and you can click any citation to confirm it against the source.