Biclay

Biclay · CiteOnly

An AI that cannot make things up.

Ask about your own documents and the answer comes back in those documents’ own words, with the document, the clause and the page it came from. No model writes the answer, so there is no fabrication to catch afterwards.

Illustration · one question, the same two documents

QuestionCan we end this contract early, and how much notice do we owe?

A general assistant

Yes. Either party may terminate the agreement for convenience with 30 days’ written notice. Early termination fees may apply if the contract is ended during the initial term.

The notice period is wrong and the fee sentence is in neither document. Nothing in the answer tells you which.

CiteOnly

“Either party may terminate this Agreement for convenience”1

Master Services Agreement Clause 14.2, page 31

“Clause 14.2 is amended by replacing ‘thirty (30) days’ with ‘ninety (90) days’”2

Amendment No. 2 Clause 3, page 2

Every word is in the file, at the place named beside it.

Two guarantees, and one honest answer

The first two hold whatever the system does on a given question.

I

Every word comes from your files

Every part of an answer is taken from the documents you supplied, word for word. If a phrase is in the answer, it is in your files.

II

Everything traces back

Each part arrives with the document and passage behind it, because that is how it was produced rather than something attached later.

III

It says so when your documents do not answer

When your documents do not answer the question, you get a clear no instead of something that merely reads like an answer. Often the clear no is what you needed.

Is this retrieval with citations?

No. Retrieval lets a model write the answer. CiteOnly does not let anything write the answer.

A retrieval system finds relevant documents and hands them to a language model, which writes the answer in its own words. The writing is where things get invented. A citation added afterwards records which document was nearby, not that it supports the sentence.

The usual remedy adds a layer: a second model grades each claim. That layer is itself a language model making a judgment. The fabrications it lets through are the ones that look most like the truth.

CiteOnly works the other way round. It locates the passages that answer your question and hands those back, each with its document and position. Nothing composes a new sentence at any point, so there is no stage to check and no answer without a source.

Two ways to answer a question about a set of documents A retrieval pipeline passes the question through a model that writes the answer freely, then a second model checks that answer; whatever the checker misses is delivered anyway. CiteOnly goes from the question to the documents straight to an answer quoted from them; the two stages in between are absent, so there is no point where new wording could be introduced. A RETRIEVAL PIPELINE Question Find passages Model writes the answer Second model checks it Answer new wording enters here errors can get past here everything it misses is delivered as an answer CITEONLY Question Your documents The answer, in the documents’ own words, with its source no writing stage nothing to check nothing was written, so nothing has to be caught
Two ways to answer a question about a set of documents A retrieval pipeline passes the question through a model that writes the answer freely, then a second model checks that answer; whatever the checker misses is delivered anyway. CiteOnly goes from the question to the documents straight to an answer quoted from them; the two stages in between are absent, so there is no point where new wording could be introduced. A RETRIEVAL PIPELINE Question Find passages Model writes the answer new wording enters Second model checks it errors can get past Answer everything it missed, delivered CITEONLY Question Your documents no writing stage nothing to check The answer, in the documents’ own words, with its source nothing was written, so nothing has to be caught
The difference is the two empty slots. Removing the free-writing stage removes the thing the checking stage exists to catch.

Already an expensive problem

Decided cases, each with a public record.

  • 2,000+

    Court decisions logged worldwide where a party relied on AI-fabricated material.1

  • A$97k

    Refunded by Deloitte to the Australian government over a report citing papers that do not exist.2

  • Liable

    Air Canada, for a fare policy its own chatbot invented.3

  • Set aside

    An Indian tribunal's insolvency ruling, after the Supreme Court found its cited judgments do not exist or do not say what was claimed.4

Run it on documents you already argue about

A pilot starts with your own files and the questions your team gets asked.