Everything it does, and where it stops.

Every answer is generated from the material you uploaded. It has no general knowledge to fall back on and is not supposed to have: the value is that it says what your policy says, not what a policy usually says.

Each answer carries a numbered reference back to the source it came from, so a visitor can check it and your team can see what somebody was told. Every page found by a site scan gets its own number rather than being folded into one.

When your content does not support an answer, it says so. This is the behavior most demos hide and the one worth asking every vendor about, because an assistant that always answers is an assistant that sometimes makes things up.

A strictness setting decides how much the assistant will attempt from partial context before declining, and how many passages it retrieves per question. Set per plan, adjustable per customer, and you can be given a bounded range to tune within rather than a free hand that can degrade your own answers.

It remembers recent exchanges within a visit, so "what about that one?" resolves. The follow-up is rewritten into a standalone question before searching, because searching on a pronoun finds nothing.

A question asked in Spanish is answered in Spanish, from the English content you uploaded. Matched to the visitor's browser automatically.

An answer with several points comes back as a real list rather than a wall of text. The set of formatting it uses is deliberately small, because this is a chat reply and not a document.

Words appear as they are generated instead of after a pause, which is the difference between a reply that feels instant and one that feels broken.

One script tag before the closing body tag. No SDK, no build step, no iframe, and it renders isolated from your page so it cannot inherit or break your styling. Title, color, corner, font and launcher text are set from your dashboard and apply live, with nothing to re-embed.

A hosted chat page at its own address, with a QR code you can print on a menu, a poster or a receipt. Put the link in a social bio. This is the option a business with an Instagram page and no site can actually use, and it works alongside the bubble if you later have both.

The same answers behind a plain HTTP endpoint, for building into your own app or piping into a system you already run.

A key can be tied to your own domains, so a key lifted from your page source will not run your assistant on somebody else's site. Keys are named, so several deployments stay tellable apart.

Planned, and listed in the product so the roadmap is visible rather than implied. They are not built, and nothing on this page should be read as saying otherwise.

PDF, Word, CSV, plain text and Markdown, including long ones. A product manual is indexed in full, so a question about page 140 is answered from page 140.

Text inside JPG, PNG, WebP and GIF images is read out automatically, so a menu that only exists as a photo, or a policy that only exists on paper, becomes answerable without anyone retyping it.

One page, or a whole site. Discovery goes through your sitemap where you publish one and follows links where you do not, respecting robots.txt, and a picker shows what was found so you choose before anything is added.

For the questions you are asked constantly, write the answer once and visitors get your wording rather than a paraphrase. Paste a policy or an announcement that never made it into a document.

Pages can be re-read daily, weekly or monthly, or on demand. Content is hashed, so a re-scan that finds nothing changed costs nothing.

In development, not available yet. The engines, schema reading and sandboxed read-only querying are built and tested; the capability is deliberately switched off, and until it ships the sources are files, pages, written answers and pasted text.

A second mode one tap from asking a question, where a visitor leaves free text instead. Automatically sentiment-tagged and surfaced alongside questions.

The assistant asks for a score out of five once per conversation, not after every reply. One tap, no submit button. Most people who would never write a sentence will still press a star, so for many businesses this is the only satisfaction signal with real volume behind it.

A visitor who would rather speak to a person leaves a name, a number and a reason, and you get a callback list with a status per row instead of a voicemail.

A visitor orders from what you actually uploaded, so the assistant cannot offer a dish that is not on the menu. Orders arrive on their own screen with a status per row, priced against your menu and totalled, and go on the moment they are placed to every destination you set up: an email ticket, a signed webhook, straight into Square's own order queue, or through Deliverect or Checkmate to Toast and most other POS systems. Pause taking orders mid-rush in one press, or hide the menu entirely.

Upload the PDF, the price list or a photo of the board, or paste the web address where your menu already lives. From a page that publishes its menu for search engines we read it exactly: sections, prices, pictures, and the sizes and extras attached to each item. Everything lands as a list you can correct: fix a price, add the dish that was missing, mark something out in one tap. Options cover sizes, extras and how it is cooked, so one item carries the choices a customer would be asked at the counter. A scanned PDF holds a picture rather than text, so those pages are uploaded as images instead.

A visitor picks a department, a day and a time in the same panel they ask questions in, and the appointment lands on your Scheduling screen. You set the departments, the hours you keep for each, how much notice you need, how far ahead you will take bookings and how many people you can see at once. Nothing can be booked outside those hours or inside that notice period, and nothing is double-booked: the time is checked again at the moment it is taken, so a slot somebody else got while this visitor was typing is refused rather than quietly overbooked. The confirmation emails both sides a calendar invitation that every calendar already accepts, so there is nothing for either of you to install.

A contact route one tap away for the questions that should not be answered by software.

Prompts on the empty chat, so a visitor who does not know what to ask has somewhere to start.

What was asked, what was answered, and how it felt, filterable and split between questions and feedback.

Unanswered questions are detected and classified automatically, with a volume trend per topic. Your documentation backlog written by the people who needed the answer, rather than guessed at in a planning meeting.

The same picture on a schedule, to the people who will not log in to look. Frequency and recipients are yours to set.

Usage and estimated model cost per account, including how much of your input is being served from cache, so "why is the bill this shape" has an answer.

An Admin changes anything. A Member sees every screen and changes nothing, which is what lets you give somebody the numbers without giving them the answers your customers receive. Enforced on the server, not just hidden in the interface.

Nobody can remove themselves, the last Admin cannot be removed or demoted, and the last person cannot be removed. None of those can be undone from inside the product, so none of them is offered.

Team size is set by your plan and pending invitations count against it, so the limit means what it says.

Each tunable setting carries an access level: hidden, visible, or editable. Editable numeric settings can be bounded to a range, because a low value is not automatically a safe one.

A question allowance, and when a month runs out the assistant stops answering new questions until it resets. No credit balance, no automatic top-up, no overage invoice.

As many windows as the traffic warrants, per visitor and per account, sliding rather than fixed so nobody gets near double the limit by straddling a boundary. A per-visitor cap sits under a much higher account ceiling, so one abusive visitor cannot lock everyone else out.

Separate from the question count, because retrieval depth and conversation memory make a question a variable unit of cost rather than a fixed one.

Abusive text is refused before any model call, so it costs nothing and never lands in your stored history.

A rate-limited visitor is told the real wait, not "try again shortly".

Each account's material is separated at every layer: storage, search and quota. Uploaded files are never given a public address, and your documents are not used to train anything.

Adding a URL resolves the hostname and checks every address it answers with, and re-checks every redirect, so neither a public name pointing at a private network nor a redirect to a cloud metadata address gets through.

A site that redirects to a sign-in page on another domain is refused rather than indexed, so one site's login page can never be stored as another site's content.

Strict transport security, no content-type sniffing, a referrer policy that keeps your dashboard addresses out of third-party logs, camera, microphone and location denied, and a content security policy. Framing is refused on signed-in screens and allowed deliberately on the embeddable chat page.

A visitor sees a plain message and a reference number. The detail goes to your admins.

Checkout names the date of the first charge, the amount, how long any introductory rate lasts and what it becomes, and asks you to confirm it. Before a free trial turns into a charge, you get an email.

Where a plan offers one you get its full feature set, at volumes set for the trial, and your card is not charged until it ends. You can decline the trial and start paying immediately, or end one early, without using it up.

Switch interval without cancelling and re-subscribing. Moving up is charged pro rata immediately; moving down is scheduled for your renewal so you are not paying twice for a period you already bought, and it can be undone before it takes effect.

Cancel from the same screen you subscribed on. You keep what you paid for until the period ends, then move to the free plan automatically.

Access reduces temporarily and restores in full when the payment succeeds, rather than the account being lost.