Introducing SupaVector

Today we are opening SupaVector to everyone.

SupaVector lets anyone build an AI that knows their own work, from their documents, their notes and their website, and put it wherever the people who need it already are. There is no onboarding call, nothing to install, and no engineer required. You describe what it is, give it what you know, and you get a link.

We built it on a simple belief. Thirty years ago, a brand without a website was a brand that could not be found. Having one stopped being a project and became a given, because the cost of not having one kept rising.

Every brand needed a website. We believe every brand will need an AI. That is the problem we are trying to solve.

What people actually want is the answer

A website is a filing cabinet. It holds everything and hands you nothing.

Somebody with a question does not want your documents. They want the sentence inside one of them. Right now the way they get it is to search your site, open three pages, skim, give up, and either send an email or leave. On your side that same question arrives by phone, by DM, by contact form, and somebody has to be awake to answer it. Nobody is awake all the time.

That gap, between the answer existing and the answer arriving, is what has quietly become unacceptable. People have spent two years being able to ask software a question and get a straight response. They now expect it everywhere, and a contact form reads as a refusal.

Closing that gap should take an afternoon, not a quarter. For most people it still takes a quarter, which is the actual reason this is not already everywhere.

Who this is for

Not developers. Or rather: not only developers.

The people who most need an AI that knows their work are the people least likely to build one. A councillor whose office answers the same forty questions every week. A clinic that fields calls about opening hours and referrals. A creator with nine years of newsletters nobody can search. A law firm whose intake is one exhausted paralegal. A shop whose sizing questions all arrive at 11pm.

SupaVector is for:

• Public figures and politicians: a place constituents and audiences can ask questions and get answers drawn from what you have actually published, instead of a contact form.
• Small businesses: hours, stock, policies, bookings, the things people call to ask when you are busy serving the person in front of you.
• Brands: product knowledge that answers at midnight, in the buyer's language, on your own domain.
• Creators: an archive that talks back. Everything you have ever published, searchable by the people who follow you.
• Government agencies: eligibility, process, paperwork. The questions that generate the queues.
• Professionals: accountants, lawyers, consultants, doctors. The screening conversation that happens before the paid conversation.
• Individuals: your own notes, journals, saved research and paperwork. Nobody else has to see it. The point is that something you wrote down eight months ago is one question away instead of thirty minutes of scrolling.

The setup is the same for all of them, and it is three steps. Tell it what it is. Give it your knowledge: files, pasted notes, or just your website address. Share the link. Most people are live in about five minutes.

Put it where your people already are

An AI that only lives on our site would be another tab nobody opens. So it goes wherever you need it:

• A link. You get one the moment you create it. Send it, pin it, put it in your bio. It works on a phone.
• Embedded on your own site. One script tag, and it opens over whatever page a visitor is already reading. They never leave.
• On your own domain. Point a subdomain at it and it is yours, with no mention of us in the address bar.

None of that is a separate project or a developer ticket. It is the same AI, reached three ways, and you can turn any of them on or off after the fact.

How fast it answers

Cost only matters if the thing is usable. A cheap answer that arrives in thirty seconds is not an answer, it is a form submission.

Looking something up is effectively instant: 90ms to check account activity, 972ms for a search across everything you know. Writing a full answer takes longer because a model is composing prose: 3.93 seconds for a single answer, 12.76 seconds for a conversational turn. All of these are wall-clock timings over six samples per operation.

We publish the slow numbers alongside the fast ones. The p95 column is the slowest one request in twenty, and we would rather you saw it here than discovered it in production.

What happens when it gets busy

An AI that is quick with ten documents and hopeless with fifty thousand is a demo, not a product.

Most retrieval systems get slower as they get fuller, because they scan more and more candidates to find the same handful of passages. We spent a long time on the policy that decides what is worth keeping close and what can be left on disk.

The figures below come from a synthetic, deterministic benchmark rather than from production traffic: a fixed-seed workload of 50,000 writes and 20,000 requests, run against the old policy and the new one.

Under that workload the wait on a busy answer fell from 5,544ms to 1,290ms, a little over four times faster, because the passages read per question fell from 3,822 to 563, an 85.3% reduction. Eight simultaneous conversations finished in 5.6 seconds against a 4.3-second serial mean, so the tenth person to ask a question is not punished for being tenth.

Full methodology, including the workload seed and the sampling, is at developer.supavector.com/performance.

What it knows, and who can see it

An AI speaking for your brand is speaking with your voice, and a wrong answer delivered confidently is worse than no answer at all.

So it answers from your material rather than from what a model happens to believe. Before it responds to anything about your business, your products or your policies, it searches what you gave it, and it is instructed not to answer from general knowledge. When the supplied material does not support an answer, it is designed to decline rather than improvise.

You decide who gets in. Keep it private while you test it. Open it to anyone with the link. Limit it to people with an email address at your company, or to a list of exact addresses. On Brand, you can restrict individual documents to specific roles. The pricing sheet your sales team sees is not the pricing sheet a visitor sees, and that filter applies to spoken answers too, where there is no visible citation to catch a mistake.

Your data stays yours. By default it lives in infrastructure we operate. On Brand you can run our gateway in your own cloud, backed by your Postgres and your model keys, so your documents, embeddings, conversations and model calls never leave that path.

Availability

SupaVector is available today.

Free gives you a real AI, not a trial: build it from your documents, notes and website, share a link that works anywhere, put your own domain on it, choose exactly who can use it, and charge for access through Stripe if you want to.

Pro, at $20 a month, adds much more usage, built-in web search, unlimited AIs, live voice conversation, and your team alongside you.

Brand, at $199 a month, adds twenty times the usage, per-role document access, your own database and model keys, prioritised indexing for online search, and customer acquisition support.

Every brand needed a website. We think the next thing every brand needs is an AI that knows its work. We would rather that be true for a two-person bakery than only for the companies who can staff an ML team.

Build yours. It takes about five minutes.