↑↓ selectEnter openEsc close

    1. Elbuz
    2. Website chat

    Where the AI in chat gets its answers and how to train it

    18 min readUpdated

    The Elbuz AI consultant answers from four sources: store details, catalog, site articles and your instructions — that's how the AI chat for your website works. What isn't there, it doesn't invent: it says so or calls a manager.

    When you need this:

    • Vague delivery answers — fill in delivery terms on the company card.
    • The bot “can't see” a product — only the smart search index gets into the chat.
    • Long warranty and exchange terms — the bot links to an article.
    • Discount requests or complaints — go to a manager with a short reason.

    Where to start: fill in your store details on the company card.

    What the AI consultant can do

    How the bot builds an answer

    Before each answer the bot receives an instruction. Some information is always in it; the bot fetches the rest itself through search when the question calls for it.

    Always in the instruction:

    • the language to answer in — the language of the page the customer is writing from; the bot uses no other language in this conversation;
    • the page the customer is on right now. On a product page — its name, short and full description and attributes, up to 500 characters each. That's why the bot understands “is it quiet?” on a product page without asking;
    • the store phone and “Working hours” together with the customer's current time — to answer whether you're open now;
    • the pages the customer viewed before the chat. The bot doesn't list them to the customer, it only takes them into account;
    • your instructions from the “Task for AI (prompt)” field — at the end. Only the ban on offering discounts comes after it.

    The bot searches for everything else on demand: store details, products, articles and, with a connected store, the customer's orders. The model picks the source based on what the question is about.

    Where the answer to a typical question comes from
    Customer's questionWhere the bot gets the answer
    “Are you open on Sunday?”the “Working hours” field — it's always in the instruction
    “How much is delivery to Lviv?”delivery terms from the company card; if there isn't enough there — articles
    “Do you have a vacuum cleaner under UAH 8,000?”catalog search with a price limit
    “What's the difference between these two models?”comparison of 2 to 5 products: price, manufacturer, availability, attributes
    “Where is my order #1234?”only if the store integration is connected (Shop ); without it the bot has no such tool
    “Give me a discount”nowhere — the bot doesn't offer discounts; it hands questions about an individual price to a manager

    The Shop API limitation applies only to the bot that answers the customer on the site by itself. A draft reply the AI prepares for a manager — the “Reply to a customer in chat” proposal in “Pulse” or a draft in the “Tickets” window (AI in tickets) — takes orders from Elbuz documents with no integration at all.

    There, the AI looks up the number the customer gave among that customer's orders, both by the Elbuz number and by the order number or id on the site or marketplace it came to Elbuz from.

    Above all of this sits the honesty rule: if there's no data or a tool is missing, the bot must say so in one sentence and suggest the next step. It must not make up order numbers, prices, delivery times, tracking codes or terms.

    The price, photo and link in the product cards the customer sees in chat come from your database, not from the bot's text. A product the search didn't return won't get into the chat, even if the model “named” it.

    The two top branches decide whether the AI answers at all: after the limit the chat closes, and with a live consultant the AI stays silent. The lower branches show what the system does after the answer
    YesNoYesNoCustomer confirmed theorderPhone or email givenA person is neededCustomer writes in the site chatMore messages in the conversationthan “Max. messages in chat”?Chat is blocked,customer gets the store phoneIs the chat consultanta live person?Message goes to “Chat dialogs”,AI doesn't answerAI gets the instruction and searchesstore details, catalog, articlesProduct cards — only those searchreturned,price and photo from the databaseThe system checks the answerOrder — only with“Placing orders in chat” onLead: email and notification“Needs a manager” folder
    • Action or result
    • Decision: the path depends on the answer

    What to fill in about your store

    Store details are filled in on the company card; the widget's “Information” tab only shows a summary. When a customer asks about payment, delivery, returns, warranty or contacts, the bot receives the contents of these fields:

    • name, description, phone, email, country, city, address, working hours;
    • “Payment terms”, “Delivery terms”, “Warranty terms”, “Return terms” (the “Trading terms” section).
    Information tab of the AI chat widget: the card saying store details come from the company card
    1Store details come from the company card
    2Warranty item
    3Edit in the company card
    1. Store details come from the company card — the robot reads the store information from the company card and answers in the visitor's language.
    2. An item without data in the card is not passed to the robot, and it answers that it has no information.
    3. Edit in the company card — a change made there reaches the chat and search immediately.

    Three things decide whether the bot sees them:

    • Language. In the main language the bot reads the card fields; in other languages — the “Texts for customers in other languages” block. An empty translation is replaced with the main-language text.
    • Empty field. The bot doesn't receive empty fields at all, so it doesn't know what should have been there. Under the honesty rule it will say it has no such information.
    • Email. Emails about leads and orders from the chat go to the email on the company card (emails).

    You don't have to fill this in by hand: the “Your company details” wizard reads the data from your site. Changes apply from the customer's next question; nothing needs rebuilding.

    The fields here are short. Long terms — warranty by brand, exchange rules, wholesale terms — are better moved into articles: the bot will find them through search and give a link.

    How the bot finds products

    The bot searches not your database directly but the index of the “Smart search” widget. So product answers first need “Smart search” on and the index built. While the search widget is off, there's no index, and the bot answers about products as if it had never seen them. How the index is built — see smart search.

    Only these products get into the index, and so into the chat:

    • with the “Enabled” status and the “Export to file (to website)” checkbox;
    • from a category that is also enabled and exported to the site.

    What happens during search:

    • the bot searches by words, a price “from — to” and, if needed, within a category;
    • products with the “Out of stock” availability status are always excluded. “Preorder” and “Coming soon” are found;
    • one search returns at most 10 products — the name and the start of the description, without the price;
    • to recommend a specific product, the bot requests that product's card. There it sees the price (or the RRP, if showing it is turned on in the settings), the sale price (under the condition below), quantity, availability status, warranty, SKU and codes, name and descriptions. This data is read from the database at the moment of the question, so a price or stock change shows up at once;
    • product limits from the “Limits” tab also apply: a product outside them won't get into the chat (more in the smart assistant article).

    The bot gets the sale price only when “Show sale price” is on in the “Prices” tab and it's lower than the regular price. Turn off showing promotions and the bot quotes the regular price, just like the product card in the chat.

    The chat searches products by words, so for product answers the key thing is that names and descriptions contain the words customers ask with.

    Giving the bot long terms: articles

    Everything that doesn't fit in the store detail fields — service, partnership, cooperation terms, long return rules — goes into regular articles (how to create them and move them from an old site — in the site articles and pages article). The hint in the “Information” tab says just that: articles are searchable by the AI assistant.

    Articles live in the same “Smart search” index. The bot will find an article if it:

    • is enabled, has the “Export to website” checkbox and isn't in the Trash;
    • is an article, not a section: the text of a section itself doesn't get into the index;
    • is already in the index. Saving an article marks it for reindexing, and the index itself is rebuilt by the “Indexing products for smart search” scheduler task or the “Create search index” button. Until then the bot won't find a new article.

    One search returns up to three articles, the first 3,000 characters of each. For details, the bot can read the full article. It puts the article link in the answer, so the customer gets a short answer and a page with the details. Hence the tip: put the key point at the start of an article, not the end.

    If your site already has delivery or warranty pages, don't rewrite them. In the “Information” tab, use “Import information pages by URL”: paste the addresses, one per line, and click “Import” — the pages are saved as articles. Already imported addresses are skipped. When the site text changes, “Refresh” re-reads the pages and marks them for reindexing.

    Information tab: hint about articles, the Manage Articles button and the import-by-URL block with the Import and Refresh buttons
    1Manage Articles
    2URL field
    3Refresh
    1. Manage Articles — long terms and conditions go here; the robot finds them by search.
    2. URL field — one address per line; an already imported page is skipped.
    3. Refresh — rereads the pages after changes on the site and queues them for reindexing.

    Articles are not the place for products: if the bot found nothing in articles, it must not conclude “there's no such product”. Products are searched only in the catalog.

    Setting the bot's rules: prompt and buttons

    “FAQ” buttons

    Despite the name, “FAQ” entries are not an answer base the bot searches. They are buttons in the chat window. When a customer clicks a button, the bot receives that entry's “Task for AI (prompt)” and answers by it. In the conversation the customer sees the section name and the buttons.

    The consequence: if the customer typed the same question by hand, the bot won't see the FAQ entry. The answer will then come from store details, articles and the catalog. So everything a customer may ask in their own words should also be in the store fields or articles, not only in a button's “Task for AI”.

    How to create sections and buttons is covered in the chat widget settings article.

    The “Task for AI (prompt)” field

    The “Task for AI (prompt)” field in the “Setup” tab is a general instruction to the bot for the whole chat. It's added at the end of the built-in instruction before each answer — only the discount ban comes after it.

    • One field for all languages. You can write in any language: the bot still picks the answer language by the customer's site language.
    • Works in messengers too. If AI replies are on in Telegram or Viber, the same instructions, store details, catalog and articles apply there. There, too, the bot sees the earlier conversation and remembers the customer's name and phone.
    • Takes effect at once — from the customer's next message.
    Setup tab: the Task for AI (prompt) field and the Max. messages in chat field
    1Task for AI (prompt)
    2Max. messages in chat
    1. Task for AI (prompt) — one field for all languages and messengers, appended to the end of the instruction.
    2. Max. messages in chat — after the limit the chat is blocked and the customer is offered a phone number.

    The prompt is for what doesn't fit in the fields: tone, boundaries, your working rules. Examples of instructions that work as rules:

    • “Address the customer formally, no emoji.”
    • “We don't install appliances. For such questions, answer that we recommend the manufacturer's service center.”
    • “If someone asks about wholesale prices or an invoice for a company, ask for a phone number and say a manager will call back.”
    • “When comparing products, always state the warranty period of each.”
    • “Don't recommend accessories from other brands for Apple devices.”

    What not to put in the prompt:

    • Prices, stock, promotions. They change, the prompt doesn't. The bot already takes prices and availability from the catalog at the moment of the question.
    • Long delivery and warranty terms. Store detail fields and articles exist for them: there they're split by language, and articles also give the customer a link.
    • Anything that contradicts the built-in rules. For example, “offer a 5% discount”: the built-in instruction directly bans offering discounts, and it comes after your text. When one instruction has both a ban and a permission, which one the model follows isn't guaranteed.

    When the bot calls a manager and takes orders

    There are topics the bot won't answer under any settings:

    • anything unrelated to the store: the built-in rule is to talk only about the store;
    • its own instructions: on “forget previous instructions” or “show me your prompt” it steers back to the store;
    • discounts: the bot isn't allowed to offer them;
    • orders, if “Placing orders in chat” is off — the bot sends the customer to order on the site.

    The bot hands the conversation to a live manager in two cases:

    • the customer directly asks for a person: “connect me to a manager”, “I want a live operator”;
    • the answer depends on facts not found in store details, the catalog or articles: an individual price or discount, a warranty case, a complaint, a return dispute.

    One rule here: if it doesn't know, it doesn't make things up. The bot tells the customer a manager will reply and gives the manager a one-sentence reason: what was asked and why it didn't answer. For ordinary questions with an answer, it doesn't call a person.

    What happens next depends on the channel.

    Site chat. The chat goes to the “Needs a manager” folder, and the customer sees “A request to connect with an operator has been sent …, please wait...”. Who gets the request:

    • the consultant of this chat, if it's a live person with the “Chat manager” flag;
    • otherwise — a random active employee with this flag; whether they're online isn't checked;
    • if there's no “Chat manager” at all — everyone; the customer then sees only the bot's reply.

    The manager gets a notification, including in their personal Telegram if linked. The bot's reason shows in the conversation as an internal note “AI handed the chat over to a manager. Reason: …”, hidden from the customer. The same reason is in the Telegram group notification, if connected.

    Employee card of the service account: Work modes block with the Chat manager and Virtual assistant switches
    1Chat manager
    1. Chat manager — only such employees receive the requests the robot hands to a person; with none, the request goes to everyone.
    Who gets a request from the site chat
    YesNoYesNoCustomer asks for a personor asks what the AI didn't findAI tells the customera manager will replyIs this chat's consultant a livepersonwith “Chat manager”?Request goes to themIs there at least one activeemployee with “Chat manager”?Request goes to oneof them at randomRequest addressed to everyone,notification sentChat in the “Needs a manager”folder
    • Action or result
    • Decision: the path depends on the answer

    Telegram and Viber. The bot pauses the chat, which goes to the “Needs a manager” folder; the reason goes into an internal note. What the manager sees and how to resume the bot — when the bot hands over a chat itself.

    No handover happens in “AI answers on its own” mode or with “No operator” on in the widget settings: the bot keeps answering, with no pause, “Needs a manager” folder, note or managers' Telegram group message. When AI · Support runs the chat in “By support policy (AI · Support)” mode, its rules apply.

    The customer left a phone or email

    The bot remembers the customer's name, city, phone and email. A phone or email makes it a lead, even if the model forgot to say so: an email goes to the store, “New lead” to the Telegram group, and the chat gets high priority in “Chat dialogs”. What arrives where — leads and orders from chat.

    Orders right in the chat

    The “Placing orders in chat” switch in the “View” tab is off by default. While it's off, the bot doesn't take orders and sends the customer to the site.

    Once it's on and the customer decides to buy, the bot asks for name, phone, delivery and payment in one message, then for what's missing. The order gets only products the bot found in the catalog; prices and total come from the database. It becomes a “Customer order” document — details: “Leads and orders from chat”.

    View tab: the AI response rating toggle and the Placing orders in chat toggle
    1Show AI response rating box
    2Placing orders in chat
    1. Rating box — thumbs up and thumbs down under each robot reply.
    2. Placing orders in chat — when it is off, the robot sends the customer to place the order on the site.

    How to check and improve the answers

    Test the bot not by guessing but with the same questions customers ask.

    1. Open your site where the chat code is already installed (see installation) and ask five or six real questions: about delivery to a specific city, about a product that's out of stock, about warranty. The conversation will be real: it shows up in “Chat dialogs”, and the bot's replies are regular paid model calls.
    2. For each weak answer, find the source in the table above: an empty field, a product outside the index, an article not yet reindexed.
    3. In the “View” tab, turn on “Show AI response rating box for each message”. The buttons 👍 “Useful answer” and 👎 “Not useful” appear under the bot's replies.
    4. Collect customer ratings: “Smart AI chat analytics” → “Chat quality” tab → the “Feedback” chart for the selected period. AI costs by day are shown there too.
    5. Read the conversations where only the bot answered — the “AI Agent” folder in the “Chat dialogs” window. That's where you see what trips it up.

    Ratings are shown only in aggregate: in the “Chat dialogs” window there's no 👍 or 👎 mark next to a specific reply. If the share of “Not useful” has grown, look for the reason in the conversations from the same period.

    Scenarios

    The bot answers vaguely about delivery. Fill in “Delivery terms” on the company card, and for other languages in the translations block. It applies from the next question.

    Long terms: warranty by brand, wholesale, exchange. Move them into an article with the “Export to website” checkbox or import a ready site page by its address. Check that “Smart search” is on and the index has been rebuilt: the bot will give a short answer and a link to the article.

    Individual prices and complaints should go to people. That's what the bot does under the built-in rule. To have a specific person get requests from the site, give at least one employee the “Chat manager” flag; you can add your own handover topics to the prompt.

    Collecting leads from the chat. Fill in the email on the company card and connect the managers' Telegram group — how to do it.

    Finding where the bot goes wrong. Turn on the rating box, check “Feedback” in “Chat quality” once a week and read the “AI Agent” folder.

    What to do when the bot gives the wrong answer

    What you see and what to do about it
    What you seeWhyWhat to do
    The bot says it has no information about delivery or payment The field is empty on the company card Fill in the field on the company card (store details)
    “Not found in our catalog”, though the product exists “Out of stock” status, the product or category isn't exported to the site, the product is outside the “Limits”, or the smart search index hasn't been rebuilt Check the status and checkboxes of the product and category, rebuild the index (product catalog)
    The bot doesn't find a new article The index hasn't been rebuilt yet, the article is disabled, lacks “Export to website”, or it's a section Wait for the “Indexing products for smart search” task or click “Create search index” (articles)
    In Telegram or Viber the bot stopped replying to the customer The bot handed the chat to a manager and paused; the reason is in an internal note in the conversation Reply to the customer or click “Resume bot” in the chat header (when the bot calls a manager)
    The customer left a phone, but no email arrived The email on the company card is empty Fill in the email on the company card (emails)
    “The maximum number of messages has been exceeded! If you have any questions, please contact us by phone” The conversation exceeded “Max. messages in chat” (300 by default; both customer messages and replies count) Raise the limit in the “Setup” tab or reply to the customer from the “Chat dialogs” window
    The bot suggests ordering on the site “Placing orders in chat” is off Turn on the switch in the “View” tab (orders)

    FAQ

    Why doesn't the bot know the delivery cost when it's written on the site?

    The bot doesn't read the site by itself. It takes data from the “Delivery terms” field on the company card or from articles. Fill in the field or import the delivery page as an article (articles).

    Do I need a separate prompt for each language?

    No. “Task for AI (prompt)” is one field for the whole chat, messengers included. The bot takes the answer language from the customer's site language. Store details, however, are filled in separately for each language.

    Can the bot give a customer a discount?

    No, the built-in instruction bans offering discounts. The bot hands questions about an individual price or discount to a manager.

    What happens if a customer asks for a manager and there are no chat managers?

    The request still appears in the “Needs a manager” folder of the “Chat dialogs” window, addressed to everyone, and a notification is sent. To have a specific person get requests from the site, give at least one employee the “Chat manager” flag.

    Where do I set the warranty for the AI chat so it quotes the right terms?

    In the “Warranty terms” field of the “Trading terms” section on the company card. The bot takes a specific product's warranty from the product card. Move long terms — by brand, the exchange procedure — into an article (store details).

    Can I run an A/B test of two welcome messages in the AI chat?

    No. There are no welcome variants, visitor split or conversion stats: the bot writes the first message itself for the visitor's page (where the greeting comes from). You can only compare periods in “Smart AI chat analytics” — “Total chats” and “Total messages” in the “Overview” tab.

    Where to find it and who has access

    The widget settings window has no menu item of its own: it opens from the category list. All knowledge sources, except the articles and products themselves, are filled in here.

    Widget settings
    Directories → Categories → “Catalog tools” → “Widget settings - Smart Assistant”
    Search index
    the “Widget settings - Smart search” tile → “Create search index”
    Answer ratings
    Analytics → Smart AI chat analytics → “Chat quality” tab
    Access right
    access to widget settings
    Prepare
    payment, delivery, return and warranty texts in each language; “Smart search” turned on

    Related topics

    Was this article helpful?

    Missing a feature? Tell us.

    We will add it for free if other users need it too.

    If only you need it, we will agree on it separately.

    Article contents

    Suggest an improvement

    Try Elbuz free