What the agent knows

The assistant answers with exactly what you gave it. No more, no less.

Ask a fresh agent “how much is delivery?” and it will honestly say it doesn't have that information. That's not a fault, it's a safeguard: an honest “I don't know” is better than a made-up price you'd then have to answer for. Add material with prices, and the answer appears.

There are three ways to teach it, and they're equally good.

Website

The shortest route. Give the address — the builder reads the pages and splits them into sources itself.

Here's my website: smak.example.com — read it and remember what it says

Where it gets pages from. From two places at once: from the sitemap (the sitemap.xml file a website makes for search engines — the builder finds it itself, including via robots.txt, nested sitemaps and compressed files) and by following links from pages. The sitemap lists pages you can't reach from the home page; links find pages the sitemap doesn't list. Each fills the other's gaps.

How many pages. By default — up to 100 pages: a small business website fits entirely. On request — up to 150 at a time; that's the cap for one website in knowledge, and it's the same for the conversation, the dashboard and schedules.

Read the website more deeply

One page. If you need just one (a new delivery condition, a separate price list), say so:

Read only this page: smak.example.com/dostavka

Add another page: smak.example.com/aktsiya

A page like this stays on its own: the builder doesn't pull in the rest of the website after it, and later re-reads only that page. The same address written differently — with a trailing slash, with “www.”, with a newsletter tag (?utm_source=…) — is the same page: it gets updated, not added a second time.

How long it takes. Up to three minutes for the deepest crawl. Ask once — asking again mid-reading won't speed anything up. If time runs out, what was read isn't lost: the builder tells you which pages are left, and when you say “finish reading the website” it starts with exactly those.

What you'll see in reply. The truth about coverage, in numbers:

Pages found: 62 (58 from the sitemap, 4 via links). Read: 60. Not added to knowledge: 2 — …/stara-aktsiya (page doesn't exist); …/kabinet (the site didn't allow access to the page).

That is: how many pages of the website it knows about and where from; how many it read; which ones didn't open and why — by name; and if the website is bigger than the cap — how many didn't fit and which ones (you can add a page you need separately).

Websites that draw their content with scripts

Some websites are “single-page apps”: the server sends an empty page, and the text only appears in the browser. Website reading doesn't have a browser, so it sees an empty space — and says exactly that: “the page content is drawn by a script in the browser”. Give the knowledge from such a website as text or files.

Website rules

The builder reads your website politely: it respects robots.txt (pages closed to robots there are not read, and it tells you what exactly is closed) and pauses between requests — longer if the website asks for it.

Knowledge from your website stays fresh

Everything read from an address — both a whole website and pages added one by one — the platform re-reads on its own once a week. The builder tells you this as soon as it has read it; the schedule shows up in “What changed”, and you can undo it like any other change.

Re-reading doesn't pull everything again. A page that hasn't changed isn't saved a second time — the platform can tell from the website's response, from the date in the sitemap, or from the text itself. Credits are spent only on what has actually changed.

What does it re-read on its own?

Re-read my website every three days

Don't re-read the website

Don't re-read this page

Re-read the website now

The interval is set in days, from 1 to 180.

A new page on the website reaches knowledge on its own — if the website was read as a whole. Nothing is added to pages that were added one by one.

A page disappeared from the website. The platform doesn't delete it: the assistant stops answering from it (so it doesn't promise customers a promotion that has ended), in the list it gets the mark “Gone from site”, and you get a notification. Then it's up to you:

Remove the missing pages

Keep it, it's still needed

If the page comes back on the website, the assistant answers from it again, and the notification closes by itself. And if the website suddenly says “not found” for most pages at once, the platform treats it as a website outage and doesn't touch anything.

You edited a page's text yourself. Your edit beats the website: re-reading doesn't overwrite it. If the page on the website changes after that, it gets the mark “Changed on site”, and again it's your call:

Take this page from the website

Keep it as I wrote it

You removed a page from knowledge — it won't come back with the next crawl of the website. Changed your mind? Name its address: “read this page”.

The same in the dashboard

In the panel on the right, tap the Knowledge row. For each page from your website it shows when it was read and when it will next be re-read, with the marks “Gone from site”, “Changed on site”, “Your edit”. Expand a row and you'll find:

• Re-read now — don't wait for the schedule (above the list there's the same button for everything at once);
• Don't re-read automatically — keep the page as it is;
• Take from site / Keep my edit — when the page changed on the website after your edit.

It also looks for an API on your website

If your website has an online store or booking system API, the builder tells you and offers to turn it into ready-made actions.

Text

Things that aren't written down anywhere: opening hours, return conditions, what to say about discounts.

Remember: we're open Mon–Sat from 9:00 to 20:00, Sunday until 17:00. The address is 12 Sichovykh Striltsiv St. Delivery in the city costs 60 UAH, free from 500 UAH.

The builder splits this into topics itself and gives each one a human-friendly name (“Opening hours and address”). One source — one topic: two sources about the same thing give contradictory answers.

To see what exactly is saved, just ask:

Show me what it says about delivery now

The builder reads a topic's text before every edit, and for good reason: an edit replaces the whole text, so anything it didn't carry over would be lost.

When your information changes, don't add a second source next to the old one — say it's a replacement:

Prices changed: delivery is now 80 UAH, free from 700. Update what was there

The builder asks for consent (the old topic will disappear) and replaces it with one up-to-date topic.

Files

The platform reads PDF, DOCX, XLSX, CSV, TXT and Markdown — up to 10 MB per file. A price list in Excel is read as a price list, not as a jumble of characters.

Give the file right in the conversation

Use the paper clip in the input field — or just drag the file onto it. One at a time: send the second one in your next message.

The dashboard asks one thing: Learn it or Send to customers. These are different things, and the file extension can't tell them apart: the first the agent reads and retells in its own words, the second it sends as is. The same PDF price list can be either. Changed your mind? Never mind, and the file goes nowhere.

The paper clip is only in the conversation with the builder. It isn't there when you reply to a customer from the “Conversations” section.

A file belongs to the agent, not the dashboard: until there's an agent, there's nowhere to put the file — and the dashboard will say so.

Good to know about files:

• a password-protected PDF won't open — save it without a password;
• a PDF made of scans (pages that are images with no text layer) won't give anything either: there's no text in it;
• a document is recognized by its file name. A re-uploaded file with the same name replaces the old one rather than being added as a second copy. A file with a different name is a different document.

Files the agent sends to customers

Files you marked “send to customers” don't go into knowledge — the assistant gives links to them in the conversation, as is. The builder knows what's there now and what to remove:

Which files can my assistant send to customers?

Remove the old May price list from there

Removing a file can't be undone, so the builder asks for consent: a link you've already given someone will stop working.

How many sources it has now

The Knowledge row in the panel on the right shows the number, and when tapped opens the list itself: what was read, from where and when. Or ask:

What does the agent know now?

The same phrase appears as a button under cards about knowledge — tap it and it goes into the input field.

The builder checks the list before adding something new — so it doesn't put a second source about the same thing next to the first.

The Knowledge row itself opens too: the sheet has the list of sources (each can be expanded, changed or removed), duplicate sources and the names the agent knows from what it has read.

Duplicate sources

The same source isn't added twice. If you asked to save the same text again (clicked twice, repeated the request because the reply was slow), the builder says such a source already exists and under what name, and doesn't create a second one. Same with a website: a page visible at two addresses (…/cli and …/cli/) is read as one.

If copies are already there — from older entries, or because the same topic was added under differently worded names — you can find and remove them. Copies aren't harmless: the agent searches all sources, and a duplicate pushes something else out of the answer.

With a button. The Knowledge row in the panel → at the bottom of the sheet, the “Duplicate sources” section → Find duplicates. The dashboard shows:

• groups of duplicates — sources with the same text; the first one is marked “Stays”, and Remove copy removes the rest;
• sources where text repeats — the same chunk appears twice; Keep text once removes the repetition and keeps the source.

The check itself changes nothing. Only what you click is removed, and one source from each group always stays. If a copy was edited while you were looking at the list, it isn't touched — and you're told why.

In words:

There are duplicates in knowledge — find them and show me

Remove the copies

The builder first names which topics are the same and where text repeats, and removes them after your word. “Undo” brings back what was removed.

If there are very many sources

Up to 500 sources are checked at a time. If there are more, or the text of one of them couldn't be read, the dashboard says so: “sources not checked: N”. In that case “No duplicates” is only true for the checked ones.

If it says “look in the “Knowledge” section”

That's the Knowledge row in the panel on the right: tap it and the list opens.

Remove what you don't need

Remove everything you read from the old website

The builder names what exactly was removed. What's removed can be brought back: the text of each source goes into the change log, and “undo” (or the “What changed — and undo” row in the panel) brings it back.

Correct or remove

Both right in the conversation:

Correct the delivery page: it's now 80 UAH in the city

Remove the promotion page, it's over

Correcting keeps the source in place and changes its text. Removing can be undone: the text is kept in the log, and “undo” brings it back. That's why the builder doesn't ask first — the rule is the same as for everything else that can be undone.

If the text doesn't save

The builder and the dashboard tell you why — and what to do depends on the reason:

• the knowledge store is overloaded — the platform has already retried the save several times. Your text is fine; in a minute the save will go through. Nothing was saved halfway;
• the text is too big for one source — retrying won't help. Split it into several topics (by section) and add each one separately;
• a failure on our side — nothing changed in knowledge; if it keeps happening, contact support.

After a failed edit the platform checks what's actually in the source and tells you exactly that: “the previous text is still in place” — or, rarely, that the topic isn't there right now and the text needs to be sent again. If the save did go through and only the reply got lost, you'll see the usual “updated” — nothing needs repeating.

Build an FAQ from what people actually asked

The platform sees what the agent didn't know, what visitors asked it in general, and which answers they themselves marked as unhelpful. Drafts are built from this:

Look at what people ask most often and put together an FAQ

Questions the agent couldn't find an answer to come first: they're exactly why you build an FAQ.

These are drafts: nothing is added until you approve. And if there aren't many questions yet, the builder will say so — an FAQ made up from nothing would be fiction, not your business.

Hide a topic for a while — without deleting it

A seasonal menu, a promotion, a temporary condition — things you need to take out of answers today and bring back in a month:

Hide the summer menu topic until winter

The topic's text is kept; the agent just stops using it. Bringing it back takes one phrase (“bring back the summer menu topic”). In the list of topics a hidden one is marked, not removed. This isn't deletion: deletion loses the text forever, and the builder asks for consent before it — but here there's nothing to ask about, because everything can be reversed.

The same works with actions: “turn off booking while we renovate” — the action disappears for the agent, while its keys and settings stay.

Names the agent knows

A product, a plan, a standard, an abbreviation, a name — in your materials these are names, but to an outsider they're ordinary words. “Kalyna” is both a plant and your tea; “Kupyna” is both a bush and the hash function from the documentation you added. So that the agent doesn't treat the question “what is Kupyna?” as off-topic and answer from general knowledge, the platform builds a list of names itself from everything the agent has read.

You don't need to do anything for this: the list is built as soon as a source is added. For things added earlier, it appears a minute or two after the agent's first conversation.

What this gives you:

• The agent knows it's its topic. It sees the most prominent names before every reply — and when it hears one of them, it searches its knowledge instead of saying “I don't know” or “that's not for me”.
• Spelling doesn't matter. “Купина”, “купиною”, “Kupyna”, “KUPYNA” — one name. A standard's number leads to it too: “ДСТУ 7564”, “DSTU 7564:2014” and just “7564”.
• It finds the exact place where the name is explained — together with the section heading, not the first chunk of text where it happens to be mentioned.

To see the list: in the panel, tap “Knowledge” — under the sources there's a “Names the agent knows” section with search. Next to each name you see its other spellings and the sentence that explains it in your materials.

When people call it something else

The platform handles letter case, word endings and alphabet on its own. But if customers use a different word that isn't in your materials, tell the builder they're the same thing:

Kupyna is the same as Купина and DSTU 7564

Remember: “Neopalyma” is what we call Kupyna

After that, a question with any of these words finds the same thing. Such spellings are marked “Added manually” in the list, and you can remove them in words too (“remove the spelling Neopalyma”). Names found in the materials themselves disappear only together with the text they're in.

If the agent “doesn't know” a word you definitely added materials about, start here:

It doesn't know what Kupyna is, even though I added it

The builder checks whether the name is in the list and tells you what's missing: the text itself, or another spelling.

Web search

Everything above is what you gave the agent yourself. But the agent can also search the internet: when the answer is neither in your materials nor in what the model knows.

Check the exchange rate before answering about prices in dollars

Check whether this carrier's shipping rules have changed

This is a built-in platform capability — you don't need to connect separate search services, and nobody asks for keys. But it's off by default: a new assistant answers only from what you gave it. To make it search, ask for it directly — the builder turns on search in the agent flow and tells you what it did.

If your installation has no search at all, the builder says so plainly instead of promising and going quiet: it knows which capabilities are actually set up.

It doesn't replace the agent's knowledge

Search brings back what's written online — and there's nothing about your business there. Opening hours, prices and terms still go in as text or via your website: otherwise the agent will answer with what it found about a similar company.

Search reaches outside, so it's slower than answering from the agent's own knowledge, and it spends credits. The agent decides itself when it's needed; if you want it never to search, just say so.

What it doesn't know

The cheapest knowledge about your own business the platform can give you isn't “customers didn't like something” but “people asked about delivery to Irpin twelve times, and we say nothing about it”.

The panel next to the conversation has a “Doesn't know N” row. Tap it to open the list of questions the agent searched for in the knowledge base and didn't find, most frequent first.

Three things worth knowing about this list:

• It's not an excerpt from the chats. A row here is the query the agent itself searched with, and it only goes to the knowledge base when it lacks facts for an answer. “Hi” and “thanks” don't end up here: the agent itself acts as the filter.
• Identical questions are merged into one. “Do you deliver to Lviv”, “delivery Lviv” and “can you deliver to Lviv?” are one row, and the number next to it says how many times it was asked. That number is what tells a one-off question apart from a gap that costs you customers every day.
• You write the answer right from the row. Type it right there — and it goes straight into knowledge as a “question — answer” pair, and the row disappears. No need to go to another section.

Rows that say the name is in what it has read, but the agent didn't find details about it are marked separately. That's a different case, and you don't need to add knowledge here — you already have materials about it. Most often the person called the thing differently from how it appears in the text: add another spelling. If the name is the same and the row keeps coming back, that's a platform bug — contact support.

Next to it there's “Not a question” — for cases when the agent searched for nonsense. A removed row comes back by itself if people ask about it again: the list describes what people are asking now, not history. A question nobody has asked for two months disappears on its own.

Digest: the same questions come to you on their own

So the list doesn't wait for you to open it, once a week the platform sends a digest — to your chat with the assistant and to your account email. By default every Monday at 9:00, and only when customers have asked something without an answer since last time: there are no empty digests.

The digest has how many questions have no answer, the five most frequent with a count, and a link to the full list; and below, one line each, everything still waiting for a person: unanswered customer requests, approval requests, the agent's questions, drafts under review, problems.

In the chat you answer questions right there. Each one has a number:

482913: Yes, we ship with Nova Poshta, 90 UAH

The answer goes into the agent's knowledge just like from a row in the list, and the question disappears. If it isn't a question —

not a question 482913

The email has no numbers: you answer questions via the link to the dashboard. Every email has “stop sending”; if the digest also comes to the chat, it stays there.

You can change this in the “What the agent doesn't know” sheet itself (the “Digest” block: don't send, weekly, daily; chat only or email only; “Send now”) or in words:

Send me the gaps every Monday by email

Digest every day at six in the evening

Don't send the digest

Send the digest now

The platform itself builds the digest from the same list and the same queues you see in the dashboard — it's not an assistant turn, and it costs no credits.

Empty doesn't mean “the knowledge base is complete”

An empty list means one of two things: the agent finds everything people ask about, or it hasn't been asked much yet. The second case is normal for a new assistant.

Building an FAQ uses these same questions: drafts are built first of all from what the agent didn't know, not from everything in a row. Or ask the builder in words:

What doesn't my bot know?

Why the agent sometimes says it doesn't know

This is a normal answer, not a failure. There are four possible reasons:

1. It didn't recognize the name as its own. If the question is one word or a name (“what is Kupyna?”), check whether it's among the names the agent knows.
2. There really is no such source — add one. (Web search doesn't save you here: your prices aren't online.)
3. The source exists, but the question is outside the agent's topic. What the agent talks about and what it doesn't do is its character, not a separate switch. Tell the builder: “let it answer questions about delivery too, that's our topic”.
4. The refusal text was changed, but the topic boundaries weren't. The refusal text only decides the wording once the agent has already decided to refuse. It won't start answering more broadly because of it.

If the agent gave a silly answer or said “not found in the knowledge base” even though the source exists, say so directly:

It gave a silly answer to the delivery question — see what happened

The builder looks at the log of the last turn and tells you the reason, instead of reconfiguring things at random.

Next

• What the agent can do — so it doesn't just talk.
• Where the agent answers — so its knowledge is finally useful to someone besides you.

Photos in the conversation

When a person sends an image in the chat — a receipt, a product photo, a screenshot of an order — the agent sees it in that same message and answers to the point. This works in the chat window on your website, in Telegram, Viber, WhatsApp, Slack direct messages and email (in Discord you can't send a photo with the /ask command); only images attached just now go into the message, up to four.

The agent doesn't read documents (PDFs, spreadsheets) itself: it says it sees the file, and an action that's given the conversation's files can read it. Same with audio and video: every uploaded file has an address, and you can pass it to any action — a transcription service, an image-processing service. The platform has no transcription of its own: you connect the one you use.