The product
Answers made of your own documents.
CiteOnly takes a set of documents you supply and answers questions about them. The answers are built from material that is already in those documents, and each part arrives with the place it came from.
QuestionCan unused annual leave be carried into next year, and is it paid out when someone leaves?
CiteOnly
“Up to five (5) days of unused annual leave may be carried into the following leave year”1
“Carried-over days must be taken by 31 March”2
Not in the documents supplied: whether carried-over leave is paid out on leaving. No passage answers it, so none is offered.
What it does
You give CiteOnly a body of documents: a contract and its amendments, a policy manual, a deal room, a set of guidelines, a case file. You ask a question in ordinary language. What comes back is the material in those documents that answers it, together with the document, the location and the surrounding passage for each piece.
Three things follow from that, and they are the reason to use it rather than a general assistant.
I
Every sentence you read is a sentence from your documents
The answer is assembled from what your documents contain, so what you read and what the file says are the same words. Verifying an answer means opening the passage named beside it, and nothing more.
II
The audit trail is always present, by design
Most systems assemble their audit trail afterwards, by logging what was retrieved and matching it back to what was written. That reconstruction is approximate, and it is where trails develop gaps.
Here the trail is produced with the answer. An answer is built out of located material, so the record of where every part came from exists before anyone asks for it, complete by design: nothing to configure, nothing to maintain, and no gap between what the system did and what the log says it did.
III
When the answer is not there, it says so
If the documents do not contain an answer, CiteOnly reports that. This is the part most teams underestimate. A general assistant asked about a clause that is not in a contract will usually produce a reasonable-sounding clause, because producing something is what it does. Getting a clear negative is often more valuable than getting an answer.
Where it fits
What it is built for, and where it stops.
-
Answers, not drafts
CiteOnly answers questions about documents you already have. Drafting a memo, summarizing into fresh prose or rewriting in a house style means producing words that are not in the source, which is exactly what it is built not to do. For those jobs you need a different tool, and probably a person.
-
Faithful always
CiteOnly can fall short on capability, never on faithfulness. It may hand you a passage that is genuinely in your documents but not the one you wanted. Because every part of an answer arrives with its source, finding that out is a matter of checking a location rather than investigating a claim. A wrong answer is cheaper to catch here than in a system that writes freely.
-
Beyond search
Keyword search returns documents and leaves the reading to you. CiteOnly answers the question and shows you the passage that answers it. The distinction matters most when the answer is spread across documents that amend one another.
Questions we get asked
Where does our data go?
CiteOnly can be self-hosted. Ask us directly and we will answer specifically.
How is this different from retrieval with citations?
Retrieval hands passages to a model that writes freely; the citation records what was nearby, not what supports the sentence. CiteOnly has no free-writing stage. The comparison is set out on the home page, and why a checking layer does not close the gap is a longer argument.
Can it work across many documents at once?
Yes, and that is where it is most useful. The awkward questions are the ones where a master document says one thing, an amendment changes it and an annex qualifies it. Each part of an answer carries its own source, so a conflict shows up as a conflict rather than being smoothed into one fluent sentence.
What does a pilot involve?
A set of your documents, the questions your team actually gets asked, and a way to tell us when an answer was not the one you wanted. That last part is the valuable one.
Which languages and formats?
Tell us what you have.
Start with documents you already argue about
The best pilot subject is the file your team keeps re-reading because the answer matters and nobody wants to be the one who got it wrong.