Ask a general AI chatbot a specific question about one of your case files and it will usually answer instantly, and sometimes it will invent a citation, a case name, or a detail that is not in any document you gave it. Grounded AI is a search-and-answer system that only drafts a response from passages retrieved out of your own uploaded documents, and it says so plainly when none of those passages contain the answer. That single behavior, refusing instead of guessing, is the difference between a tool you can hand to a document review and one you have to double-check line by line.
This article looks at the mechanism behind that refusal: why general chatbots hallucinate in the first place, how a grounded system decides there is nothing to say, and what an honest not-found answer actually protects you from.
What Makes an AI Answer "Grounded" Instead of Guessed?
A grounded answer is built only from text retrieved out of a specific source, with a citation pointing back to it. A guessed answer is built from the model's general training, with nothing behind it but pattern-matching. The retrieval step is what makes the difference checkable: you can open the cited document, go to the cited page, and confirm the sentence is actually there.
Vorticel's retrieval step works the same way for every question: your question is embedded, compared against the embeddings of your firm's own document passages, and the closest matches, and only those, are handed to the model to draft from. Nothing outside that retrieved set reaches the answer. Vorticel sends extracted document text to Voyage AI to create those embeddings and sends only the retrieved passages to Anthropic to draft the answer; the original file itself never leaves encrypted storage at any point in that process.
Why Do General Chatbots Hallucinate Citations?
A general chatbot answers from patterns learned across its training data, not from your documents, so when it produces a case name, a page number, or a quoted clause, it is predicting what a plausible answer looks like, not looking one up. If the real answer is not in its training data, or your question is specific to a document it has never seen, it will still produce a fluent, confident-sounding response, because fluency is what it was trained to produce.
That is the mechanism, not a defect that gets patched away with a better prompt. Legal work has already seen the real cost of this: attorneys have been sanctioned by courts after filing briefs that cited cases a general chatbot invented outright, because nothing in that chatbot's process ever checked the citation against a real source before it was written down. A tool with no retrieval step has no source to check against, so a confident wrong answer and a confident right answer look identical from the outside.
This is why the fix is not a better prompt or a politer disclaimer. Prompting a general chatbot to only cite real cases does not give it a source to retrieve from; it only changes the wording of the same guess. The fix has to happen earlier, in the architecture: retrieve first, from a known set of documents, then draft only from what was retrieved. A tool with no retrieval step has no source to check against, so a confident wrong answer and a confident right answer look identical from the outside. Our earlier comparison of grounded AI and general chatbots walks through this gap in more depth.
How Does a Grounded System Decide When to Refuse?
It refuses when nothing retrieved is close enough to the question to count as an answer. Every passage in your document library has already been converted into a numeric embedding, a representation of its meaning rather than its exact wording, as covered in how vector embeddings power legal AI search. Your question gets the same treatment, and the system measures how close your question's embedding sits to each passage's embedding.
When the closest passages fall outside a set similarity threshold, there is no passage worth drafting an answer from, and Vorticel returns an honest statement that your documents don't contain an answer to this, instead of falling back on general knowledge. That threshold is a deliberate design choice, not a bug: a system tuned to always produce an answer will eventually produce a wrong one with the same confident tone as a right one.
Step by step, a refusal happens like this
- Your question is converted into the same kind of embedding as every document passage.
- The system searches your firm's own passages for the closest matches, never another firm's documents.
- Each match is scored for how close it actually is to your question, not just whether it shares keywords.
- If the closest matches fall outside the threshold, no passage is handed to the model to draft from.
- You get a plain statement that the documents you uploaded don't contain the answer, instead of a guess.
The Citation Is the Proof, Not Just a Courtesy
A citation to a document name and page number is not decoration on top of the answer, it is the entire reason the answer can be trusted without re-reading the source yourself. Every answer Vorticel drafts points back to the specific document and page a passage came from, so checking a grounded answer takes seconds: open the document, go to the page, confirm the sentence.
Manually re-verifying every AI-drafted fact against the original file is exactly the kind of task that eats an afternoon on a document-heavy matter, cross-checking a deposition summary, a lease abstract, or a contract comparison sentence by sentence against the source file. Vorticel's retrieval and citation pipeline is built to make that check trivial instead of tedious, because the citation is generated at the same moment as the answer, from the same retrieved passage; it is never added after the fact.
Grounded Refusal vs. a Confident Guess
| Situation | General chatbot | Grounded, cited AI |
|---|---|---|
| Answer not in your documents | Produces a fluent guess anyway | States plainly that the documents don't contain an answer |
| Case name or clause cited | May be invented from training patterns | Points to a specific document and page you can open |
| Checking the work | Requires re-researching the claim from scratch | Requires opening one cited page |
| Confidence tone when wrong | Identical to when it's right | Refuses instead of sounding equally confident |
Doesn't Refusing More Often Make the Tool Less Useful?
It can feel that way the first time a question comes back as not found instead of an answer. The honest trade-off is this: a tool that always answers is not more capable, it is only more willing to guess, and in document review, contract comparison, or deposition prep a wrong guess costs far more than a not-found costs in follow-up. A grounded refusal tells you something real, that the answer genuinely is not in what you uploaded yet, which is often useful information on its own, pointing you to what document is still missing from the file. On a small-firm team without a research department to catch a bad citation before it reaches a filing, that early warning matters more, not less: the tool telling you what it doesn't know is doing part of the review work a second associate would otherwise have to do by hand.
Vorticel does not use your documents, questions, or answers to train any model, and per-firm libsodium encryption plus tenant isolation by firm keep the underlying files separate from every other client on the platform; see the security page for the full data-flow breakdown. Honest refusal and tenant isolation come from the same design principle: don't produce output, or grant access, that the evidence doesn't support.
See a Grounded Refusal for Yourself
The fastest way to judge this is to ask a real question against a real document. Start the 14-day free trial, no card required, upload one file, ask a question it answers well, then ask one it genuinely can't, and read the two answers side by side. Start your free trial and see what an honest not-found answer looks like next to a cited one.