The document store behind your platform.
Every attachment in every file, read page by page rather than sampled. At one customer that is 44,956 documents and 127 GB, and coverage is 100% of everything the document system will actually serve.
An answer with no source behind it is not an answer.
Every claim Mayeshal makes names the document and the page it came from, and the link opens that file at that page. This is the feature a firm asked for by describing its absence: their own words for the problem were that the AI could not point at the page it got an answer from.
Shipped and running in production today
331,486 pages indexed at one customer
The link opens the file at the page
The problem
A confident paragraph is not evidence.
A general assistant will give you a fluent answer about a file it has read a fragment of, and the answer will be right often enough that people stop checking. That is the dangerous state. The first time it is wrong in a way that matters, nobody can reconstruct where the sentence came from, because there is nothing to reconstruct.
In a regulated business the requirement is stronger than accuracy. You have to be able to show where an answer came from, months later, to someone who was not in the room. Across five separate products there is no single trail to follow, and inside a product with no citation there is no trail at all.
So the design rule here is blunt. If the pages do not support the sentence, the sentence does not get written.
How it works
Index the page, cite the page, open the page.
Citation is not a formatting step at the end. It is a property of how the corpus was built in the first place.
- 01
Every page is opened once
A text layer where there is one, optical character recognition where there is not, and a vision model for photographs and screenshots. The status of every page records how it was read, so a recovered rendering is never mistaken for the original.
- 02
The passage is located, not paraphrased
A citation returns the file, page N of M, and the matching passage itself. You are not trusting a summary of a summary, you are looking at the sentence the model looked at.
- 03
The link opens at that page
Clicking a citation opens the document at the cited page rather than at page one of a nine thousand page binder. On one live matter a phrase search returned twelve page level hits across 9,339 pages.
- 04
The gap is named instead of filled
Where the pages do not support a conclusion, the tool says so and names what is missing. The instruction it runs under is explicit about not filling the gap from memory.
What it connects to
Where the pages come from.
Citations are only as good as the corpus underneath them, so the corpus is the product.
Your own storage, after first read.
Once a document has been read, a copy sits in your tenant's own storage. Opening it takes hundredths of a second, and more usefully it still opens when the source system is throttled.
The record around the document.
A citation is more useful when the system also knows whose file the page belongs to, which is what the mirror of your records is for.
Mail and meeting transcription.
Threads and transcripts become citable objects alongside documents. These connectors are authorised in your own tenant during implementation.
Where it matters most
Anywhere somebody may have to prove it later.
The industries where a cited page is not a nicety, and the ones where it quietly saves an afternoon.
Legal
Records, deadlines, negotiations and the document file never sit in one place, and the client is asking one question about all of it.
Legal pack, live in productionFinancial services
Client records, a books and records obligation, and an examiner who wants to see what you knew and when.
Pack built and signedAccounting and tax
Client documents arriving as PDFs and email, and a person moving them into a spreadsheet by hand.
First customer build
Said plainly
What a citation is, and what it is not.
A citation points at your own documents. It is not a legal authority and it is not a published source. Case law search and citation verification against published authority is a separate capability with its own page.
A page that cannot be read confidently is flagged rather than guessed at, and the page status records how it was read. That is deliberate: a confident wrong reading inside a client file is worse than a plain statement that a person needs to look.
And it does not follow that every question has an answer in the file. Often the correct output is that the record does not support the conclusion and here is the document that would. That is the system working, not failing.
Get started
See it read one of your own matters.
Thirty minutes, on a screen share, on a matter you pick. We connect nothing, install nothing and ask for no credentials to have that conversation. If it is not obviously useful in the first ten minutes, we will tell you so and you get your afternoon back.