↑↓ selectEnter openEsc close

    1. Elbuz

    How to train the configurator AI assistant on your catalog

    8 min readUpdated

    The configurator chat assistant answers from your products, not with generic phrases: Elbuz builds a search index of your catalog and sends only changed products on updates.

    When you need this:

    • A customer asks which motherboard fits a CPU — the assistant searches products in the build.
    • You updated the catalog — a rebuild skips and doesn't bill unchanged products.
    • A build was stopped halfway — what was written stays; the next run adds the rest.
    • You want to see what the configurator lacks — “Chat history” shows questions the AI couldn't answer.

    Where to start: turn on “AI training” in the configurator card, then run the index build.

    What it does

    • AI chat in the configurator — an assistant with your prompt, model, message limit and frequently asked questions.
    • AI product index — the assistant finds answers among the products in the configurator structure.
    • A rebuild sends only new and changed products.
    • Stopping keeps what has already been written; the next run continues.
    • Chat history shows what customers ask about.

    How to turn on the chat and AI training

    In the configurator card, on the “Setup” tab, the “AI chat settings” section:

    • “Show AI chat” — shows the chat to customers;
    • “AI training” — extended training of the assistant on your catalog; without it, the index isn't built;
    • “Max. messages in chat”;
    • “List of frequently asked questions”;
    • “AI Model” and “Task for AI (prompt)”.

    The chat greeting goes on the “Information” tab — the “Welcome message in chat” field for each language.

    For the chat to answer based on products, a search index is built for them — the “AI product index” tile. After major catalog changes, rebuild the index.

    How to build the AI product index

    The “AI product index” tile (tooltip “Create index for AI”) is in the “Categories” window, next to the configurator tile. It opens a separate “Create index for AI” window with the “Settings” and “Log” tabs; the “Run” button starts the build, and the window then switches to the log on its own.

    The “Create index for AI” window on the Settings tab
    1Run
    2Force reindexing of all records
    3Category ID
    4Log tab
    1. Run — starts building the index and switches the window to the log.
    2. Force reindexing of all records — erases the index of all products in the account.
    3. Category ID — adds a category with its subcategories to the categories from the configurator structure.
    4. Log tab — shows the progress and the result of the build.

    What happens after “Run”:

    • If no active configurator, assistant or smart search has “AI training” turned on, the build doesn't start at all and nothing appears in the log. Turn on the toggle and save the card first.
    • Only active products with the export flag that sit in active categories with the same flag are indexed. A product without a name is skipped.
    • The build takes the categories from the “Structure” of active configurators with “AI training” on, including subcategories; the category in the “Category ID” field is added to them. If those configurators have no categories in their structure, the whole catalog is indexed, or, when “Category ID” is filled in, only that category with its subcategories.
    • For each product, a text is assembled from its name, category, SKU, descriptions and attributes. If this text hasn't changed since the last build, the product is skipped and costs nothing; the log counts such products as cached.
    • Changed and new products are sent to OpenAI, and each one is written to the index separately as soon as its response arrives. This is what gets billed: each product that gets a response is a separate charge based on the number of tokens in its text.

    The “Force reindexing of all records” checkbox first erases the index of all products in the account, not just the selected category. All products included in this build are then sent and paid for again, while the rest have an empty index until a build covers them.

    You can see that the build went through on the “Log” tab: at the end, the line “Indexing completed in ‹time›, total records ‹N› (‹M› cached)” appears, and along the way there are interim lines “Processed records: ‹number› (‹total›)” at every 500th product, if it had to be sent.

    Products are sent up to 30 requests at a time; the program waits up to 15 seconds for each product's response. If 5,000 requests add up within a minute, the build pauses and logs “Waiting ‹N› sec until the next run”. There's no cap on the number of products in a build.

    The top two branches fire once per build, then the loop goes through the products: only the “Product goes to OpenAI” branch is billed, and “Stop” is checked before each next product
    noyesyesnoyesnonoyesyesno“Run” buttonDoes anything have“AI training” on?The build doesn't start“Force reindexingof all records”?Index of all products erased“Stop” clicked?“Operation canceled!”Product text changedsince the last build?Skipped, counted as cachedProduct goes to OpenAIresponse written to the indexMore products?“Indexing completed…”
    • Start or end
    • Action or result
    • Decision: the path depends on the answer

    What happens if you stop the build

    The “Stop” cross button on the “Log” tab ends the build before the next product: “Operation canceled!” appears in the log, and there's no “Indexing completed…” line.

    • The exception is stopping after the program has already gone through all products and is waiting for the last responses: in that case the “Indexing completed…” line still appears.
    • Responses that arrive after the stop aren't written to the index. Whatever was written before the stop stays in the index.
    • A product that got no response keeps its previous index (or has none, if it didn't have one yet or force reindexing erased it).

    Running it again without the “Force reindexing of all records” checkbox is safe: it skips already indexed products whose text hasn't changed, doesn't charge for them again, and picks up the rest.

    When the index updates on its own

    Editing the configurator card doesn't update the index. Saving a product card updates that product's index only when smart search is on and isn't marked for full reindexing. Otherwise, changes to descriptions and attributes reach the index with the next build or smart search reindex.

    What to look for in chat history

    Customers' conversations with the assistant are saved on the “Chat history” tab of the configurator card. Questions the AI couldn't answer reveal gaps in attributes and in the structure — add the missing attributes or steps and rebuild the index.

    Scenarios

    The chat assistant should answer based on the products in the build

    In the configurator card, turn on “AI training” and save, then in the “Categories” window open the “AI product index” tile and click “Run”. The index gets products from the categories in the “Structure”, with subcategories. What won't happen: with no categories in the structure, the whole catalog is indexed, and later configurator card edits don't update the index.

    You updated the catalog and need to refresh the index without paying for everything

    In the “Create index for AI” window, click “Run” without selecting the “Force reindexing of all records” checkbox. Only products whose text changed since the last build are sent and billed; the log counts the rest as cached.

    What to do when the chat gives the wrong answers

    What you seeWhyWhat to do
    The log is empty after “Run”no active configurator, assistant or smart search has “AI training” turned onturn on the toggle in the card, save it and run the build again
    The AI chat answers in generic termsthe catalog index hasn't been builtturn on AI training and run the index build
    The chat doesn't know about new product descriptionsthe product edit didn't reach the indexrun the build without the force reindexing checkbox
    The log has no “Indexing completed…” linethe build was stoppedrun it again without the checkbox — it picks up the rest

    FAQ

    What does training the AI on the catalog give me?

    The chat assistant starts answering based on your products, not with generic phrases. For this, a search index is built; after major catalog changes, rebuild it.

    Why did the whole catalog get into the index, not just the products in the build?

    The “Structure” of configurators with “AI training” on has no categories. Add steps or enter a “Category ID” in the index window.

    Will I pay again if I run the build once more?

    No, not for unchanged products: they count as cached. All products in the build are billed again only with the “Force reindexing of all records” checkbox.

    Why did “Waiting … sec until the next run” appear in the log?

    5,000 requests added up within a minute, and the build pauses. After the pause it continues on its own; there's no cap on the number of products.

    How long does it take to get a response for one product?

    Up to 15 seconds; products are sent up to 30 requests at a time. A product without a response keeps its previous index.

    Where to find it and who has access

    Menu
    Directories → Categories → “Widget settings — Configurator” tile → “Setup” tab
    Index
    Directories → Categories → “AI product index” tile
    Access right
    the “Categories” window in “Group permissions”
    Prepare
    a configurator with a completed structure

    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