An answer desk that never makes things up

An owned licence is $12,000 once plus $2,400 a year for updates and support, or $6,000 a year all-in. Hosted starts at $149 a month and never meters your questions. It costs more than a generic chatbot because it is a different thing: you own the whole system, your content never goes to an outside AI service, and there is no per-question meter behind it. If your budget is a cheap chatbot, we will tell you honestly that this is not a fit.

Ask yours a question your policy answers plainly, then ask it something your site has never covered. Most will answer the second one anyway, confidently, with no source to check. Viavitna deletes claims it cannot tie to your text and declines when your content does not cover the question. The honest test is to run ten of your real questions through both, which is exactly what the no-charge content-fit review is.

With an owned licence, nowhere. The engine, the AI models, and your content all sit on a machine inside your network, and neither your material nor your visitors' questions ever leave it. There is no outside AI service anywhere in the answer path in any configuration. On the hosted plan your content sits on infrastructure we operate — that is the one setup where it leaves your building, and we consider it the lesser choice.

Less than people expect, because there is nothing to rewrite or re-tag. Web pages, HTML, Markdown, and plain text go in as they are, and we can crawl your live site to build the first draft for you. Anything in another format has to be turned into text first, and the content-fit review is where we find out how much of that there is. Adding it to your website afterwards is a single line pasted once. The real work, when there is any, is fixing content the draft shows is unclear.

A perpetual licence keeps working on the version you have — that is the honest behaviour for something you bought. Maintenance keeps the engine and the models current; letting it lapse stops the updates, not the assistant. We will also tell you the part vendors usually bury: the licence is enforceable, and a licence marked revoked stops the box answering, though a licence service we cannot reach never stops a paying customer. On hosted, export and wind-down terms are agreed before service starts, and moving from hosted to an owned licence later is a supported path we would rather you took.

It is running today at viavitna.com, answering live questions with citations from its own published pages, and you can go interrogate it right now. Its deployments, security review, and acceptance runs are dated and written down, including one release switch that failed closed on a permissions fault, rolled itself back, and passed on the second attempt. The owned image has been built, encrypted, licence-unlocked, and proven answering end to end; yours would be built to current source when you order. We would rather show you the record than tell you it is mature.

Your visitors keep getting answers, just plainer ones. When the writing model is unavailable, Viavitna quotes the relevant passage of your own content directly, still cited, or says plainly that the material does not cover the question. It is a downgrade in polish, never a downgrade in truthfulness, and it is the same floor that catches every other failure.

Yes, and this is one of the reasons people buy it. Every answer carries a receipt binding which version of your content was serving and which passages were used, and for any period a report can be pulled covering what was published, from which sources, how many questions came in, and every time it declined to answer. That report is produced without ever storing the words a visitor typed, and it proves what the system decided — not that your source material was itself correct.

Your content is treated as reference material to quote, never as instructions to follow, and unsafe requests or attempts to fish for credentials are refused before anything else runs. The protection that actually matters is structural rather than a list of blocked phrases: even if an injected instruction were followed, the resulting text would be uncited and unsupported, and the verification step deletes exactly that. The pre-screen phrase patterns are written for English and Spanish; the structural check does not care what language anything is in.

Because nothing you upload is public until you decide it is. Your material becomes a private draft that visitors cannot reach, and you spend as long as you like asking it the questions your team hears every week and reading the citations. Publishing is a separate, deliberate click, every earlier version is kept and hash-checked, and bringing an old one back puts it in review rather than on your site.

We would rather you hear this from us than discover it later. As of 2026-09-02:

• It can only say what your content says. If your policies are vague, contradictory, or out of date, Viavitna will abstain rather than paper over the gap — useful, but it means content work is often the real first project.
• Content has to arrive as web pages, HTML, Markdown, or plain text. There is no document or scanned-file reader: anything in another format has to be converted to text first, which we sort out during the content-fit review rather than after you have bought.
• It does not look anything up. It cannot check an order status, an account balance, or a case number, because it has no connection to your records. Anything personalised needs a separate, scoped integration.
• It is not an open-ended creative assistant. Ask it something outside your material and it declines. Buyers who want a chatty general assistant should buy a different product.
• The list of questions it could not answer shows counts, languages, and the sources that came closest — never the words a visitor typed. That is the deliberate trade: you can find the gap, but you cannot read the message.
• The smaller model that runs on the owned image sometimes declines a question it could have answered, particularly around relative dates and time phrasing. The result is over-caution, never invention.
• The pre-retrieval refusal patterns for unsafe or account-lookup requests are written for English and Spanish phrasing. The structural protection — deleting any claim your own text does not support — is language-independent and is what actually carries the promise.
• Answers are only as fresh as the version you published. The drift check compares content that has a live public page behind it; a page whose text is assembled in the visitor's browser is reported as unverifiable rather than guessed at, and nothing is ever republished automatically.
• The owned licence is enforced: the disk is encrypted and your key unlocks it, and the box can be configured to check its licence status with us, which means a licence marked revoked stops it answering. An unreachable licence service never stops a paying customer, but you should know the mechanism exists before you buy.
• The pre-launch security review and the ongoing quality gates are ours, run in-house and documented, not a certificate from an outside firm. We have not commissioned a third-party penetration test or a formal certification.