↑↓ selectEnter openEsc close

    1. Elbuz

    How to run AI description generation for a catalog or a category

    11 min readUpdated

    Elbuz queues a job for every checked field and writes the model's answers straight into the cards — in the background, while you work on something else.

    When you need this:

    • A catalog without descriptions — a run on the whole catalog fills only empty fields.
    • A category has its own template — launching from that category uses its prompts.
    • One description needs rewriting — “Generate texts (AI)” in the product row menu.
    • You need to know the price in advance — the line under the model list estimates the cost of a thousand descriptions.

    Where to start: a template in “Text generator” → “Start generation”.

    What bulk generation does

    Which scope to run with

    The scope isn't set in the launch window — it depends on where you opened the window from. There are three working options:

    The whole catalog

    For the whole catalog, prompts come from the universal template. What gets processed depends on the entry point:

    • “Start generation” in the universal template — the fullest option: the window shows all four blocks, and every product, category, manufacturer and article with an empty checked field is processed.
    • The tile in the “Categories” window — also the whole catalog, but categories only.
    • “AI & content” → “Bulk text generation using AI” in the base catalog — products, taking the category in the tree into account: the root or nothing selected means the whole catalog; a specific category shows “Category selected” in the header, the run is limited to it, and prompts come from its template.

    One category

    This is how the window opens from a specific category's template with the “Start generation” button, and from the category card with the “AI generation” button. The window header shows “Category selected” with the category name. Prompts come from that category's template, and anything missing there comes from the universal one.

    The two entry points differ. The template shows all blocks and by default covers products directly in this category. The category card shows only the “Categories” block: it writes the fields of the category itself — its description, , meta tags — and leaves its products alone. In both cases “Including subcategories” extends the selection to the whole branch.

    One product

    The “Generate texts (AI)” item in the base catalog row menu opens the same window with “Selected product” in the header — only that product is generated. Or, without leaving the launch window: the entity block header has a small field — enter an item's number there to limit the run to that item.

    How to start a run

    Elbuz writes the text straight into the card field, and the previous value isn't saved anywhere. That's why the “Only update empty fields” checkbox in the launch window is on — clear it only on purpose. Every model request is charged to the account balance.

    The “Bulk text generation using AI” window has three tabs: “Settings”, “Log” and “Statistics”. The window itself prompts the order of steps: choose a model, check the fields, click “Run”, watch progress in the “Log”.

    AI model. The list offers three options: GPT-6-Sol, GPT-6-Luna and Gemini 3.8 Flash.

    • By default it shows the model set in the account settings; if none is set there, the first one in the list is selected — GPT-6-Sol, the most expensive.
    • The cheapest in the list is GPT-6-Luna.
    • If you've added your own AI service key, it becomes the first item in the list and is the one selected.
    • Under the list there's a line with an approximate cost — it's recalculated each time you change the model (what it costs).

    Two checkboxes. “Only update empty fields” is on by default and means exactly what it says: only items whose checked field is empty are processed. “Including subcategories” is off and only matters when the window is opened for a category.

    Fields. Below come blocks by entity, each with field toggles. Four toggles are already checked: Full description, HTML H1 tag, HTML title tag and Description meta tag. This is the starting set, and you can change it.

    The “Select all” and “Clear all” links are both above the blocks and in the header of each block. If no field is checked, “Run” responds with the message “Check at least one field to generate”.

    The “Products” block header has two more numbers: FAQ and RV — how many questions and how many reviews to ask the model for. Both are empty. A number entered here takes priority; an empty field means “take the number from the template” (where it comes from).

    Settings tab of the bulk AI text generation window
    1Run
    2Job queue
    3AI model
    4Only update empty fields
    5Cost estimate
    1. Run — puts the generation jobs in the queue.
    2. Job queue — shows what has already been queued.
    3. AI model — the choice affects the price.
    4. Only update empty fields — on by default, protects filled fields.
    5. Cost estimate — recalculated when the model changes.
    Products block with field chips and the header of the next block
    1Select all / Clear all
    2ID
    3FAQ and RV counts
    4Fields checked by default
    1. Select all and Clear all — work only on this block.
    2. ID — limits the run to a single item.
    3. FAQ and RV — how many questions and reviews to request from the model.
    4. Checked chips — Full description, HTML H1 tag, HTML title tag and Description meta tag are on by default.

    What language the text comes out in

    The launch window has no language choice. The output language is the main catalog language: the one marked as default in Directories → “Languages”. Prompts are written in the same language, and the window says so directly: “Prompts are written in the main language (…) — other languages are filled by automatic translation”.

    The generator writes only in the main language of the “Languages” directory, no matter how many other languages are enabled. You can't switch this in the launch window: change the main language in the “Languages” directory first, and only then start the run.

    For non-main languages, Elbuz creates empty description rows before the run, so a translator has somewhere to write. The generator doesn't translate anything itself.

    What it costs and where to see the spending

    The bill is made up of individual requests: one request for each field of each item. The price of a request depends on the amount of text — the unit here is the , and both the prompt and the answer are counted.

    So a Title costs many times less than a full description, and the total is multiplied by the number of products.

    The line under the model list puts this in plain numbers: “1000 product descriptions will cost €… if 1 description is 1000 characters long.” For GPT-6-Luna it shows €2, for Gemini 3.8 Flash — €30, for GPT-6-Sol — €40. The cheapest and most expensive models differ twentyfold — the main reason to check the model list before the run.

    Дані таблицею
    ПоказникGPT-6-SolGemini 3.8 FlashGPT-6-Luna
    A thousand descriptions of 1000 characters40302
    The cost of a thousand 1000-character descriptions with the three models in the list. GPT-6-Sol costs twenty times more than GPT-6-Luna — the chart makes it obvious.

    Spending is charged to the account balance. Each request becomes a separate row in the “Operation log” window (Account → Operation log) — it has a “Cost” column, and next to it you can see which field was written for which item. Log rows are kept for 14 days.

    You can work with your own AI service key — add it with the “Add API access key” button in the same window.

    One detail for owners of their own key: in the model list that item is marked with an asterisk and the words “my key”, and the cost line under it always shows the same number — €5.67 per thousand descriptions — whatever model runs behind the key. Go by your AI service's price list, not by this line.

    What happens after you start

    Each checked field of each entity becomes a separate queue job. So a run of “products, four fields” is four jobs, not one. You can close the window: the work goes on in the background and isn't affected.

    Jobs run in parallel, in several threads. For most fields the limit is forty simultaneous requests, for the full description — ten, for the short one — fifteen, for FAQ and reviews — ten each. The limits differ because AI services count a tokens-per-minute limit: long answers hit it faster than short ones.

    The “Log” tab shows a live stream: a row for every finished field with the item name. The “Stop” button removes everything from the queue that hasn't started yet; whatever is already in progress gets finished. When no unfinished items are left, the log shows “Text generation finished”.

    Two buttons in the panel lead to neighboring windows: “Job Queue” — to see the generation jobs as a list, and “Operation log” — to see the spending.

    The text goes into the card as is, without review or approval: the model can invent a property the product doesn't have, so read the first cards. Automatic accuracy checking (“Copy fact-check”) exists only in the landing page generator.

    If someone fills in the field by hand between the moment the job is queued and the moment it runs, the model's answer isn't written: the “field is empty” check is repeated at the moment of saving. The request has already been made and paid for, though.

    Limits of a single run

    There's no cap on the number of items. Every product, category, manufacturer or article that falls within the scope and passes the “Only update empty fields” condition gets a job, however many there are. Elbuz doesn't estimate in advance how long the run will take.

    How the queue works through jobs:

    • “Run” starts queue processing itself, without waiting for the schedule. After that the queue picks up the work on schedule every minute.
    • In one pass the queue takes as many items as the field's thread limit allows — 40, 15 or 10 — and sends requests with a one-second pause between them. The next pass takes the rest.
    • If generation has used more than 100,000 tokens in the last three minutes, the limit drops to five simultaneous requests.
    • If the account has several keys, the thread limits are multiplied by the number of keys and divided by 1.35.
    • Elbuz waits up to 420 seconds for the model's answer. An item that hasn't finished in 10 minutes is considered stuck and is put back into work. A retry after an AI service error also counts as an attempt. After the fifth attempt the item is no longer taken: its field stays unchanged, and the run continues to the end.
    • Every 10 seconds the launch window checks whether any unfinished items are left. When there are none, the run is considered complete.
    • Clicking “Run” again for the same field and the same scope doesn't add a second job — it replaces the unfinished previous one, so unprocessed items aren't duplicated.
    • Jobs queued more than three days ago are deleted as soon as any new job is queued — together with whatever in them hasn't been done yet.

    A single-item run doesn't use the queue. If generation is limited to one product (the “Generate texts (AI)” item or the ID field), to one manufacturer or article via the ID field, or, for the “Categories” block, to one category without the “Including subcategories” checkbox, requests are sent one after another right while “Run” is being processed, without parallel threads.

    The “Stop” button works here too: the next items and fields are no longer requested, and only the request in progress at that moment is finished.

    What happens after “Run”. The first fork is AI access, the second decides whether the run goes through the queue. The bottom node is shared by both paths: this is where the “Only update empty fields” checkbox is checked a second time, at the moment of writing.
    noyesyesnoyesnoClicked “Run”AI features availablein this account?Refusal with the reason explainedno jobs createdSingle-item run?ID field, one productor the “Categories” block for onecategoryRequests are sent one after anotherwhile “Run” is processedA queue job for each fieldan identical unfinished one isreplacedA queue pass takes up to 40, 15 or10 items1-second pause between requestsThe answer is written to the fieldwith “Only update empty fields” —only to an empty oneAny unfinished itemsleft in the queue?The log shows “Text generationfinished”
    • Action or result
    • Decision: the path depends on the answer

    Scenarios

    First run: one category, one field

    Don't make the first run on the whole catalog and all fields at once. Take one category and one field — “Full description” — and look at a dozen results in the cards. A prompt takes a minute to fix, while regenerating a thousand descriptions costs as much as generating them the first time.

    What to do. Open the launch from the category template, click “Clear all” and check only “Full description”. Model — GPT-6-Luna, the “Only update empty fields” checkbox — on.

    Products only, no blog articles

    What to do. In the launch window opened from a template, click “Clear all” in the “Articles” block. Or open the launch from the base catalog: there only the “Products” block is visible.

    What you get. Requests go only for products, and you don't pay for articles.

    Rewrite one product's description

    What to turn on. In the base catalog, in the product row menu — “Generate texts (AI)”. Clear the “Only update empty fields” checkbox and leave checked only the field you need to rewrite.

    What you get. The window header shows “Selected product”. Requests go one by one while “Run” is processed, without the job queue, and the new text replaces the old. What you won't get: the previous text isn't saved, so it can't be restored. After “Stop”, the remaining checked fields aren't generated; the field already in progress is finished.

    What to do when a run doesn't start or the text is wrong

    What you seeWhyWhat to do
    “No AI access: AI service API key is not configured.” There's neither a platform key nor the account's own key. Add your own key with the “Add API access key” button (about keys)
    “Check at least one field to generate” You clicked “Run” without checking any toggle Check the fields you need or click “Select all” in the block you need (starting a run)
    “Our servers are currently overloaded. Please try again later.” The AI service is overloaded. The most common generation error on the service Elbuz retries such a request once on its own. If the retry fails too: in a queue run the item goes back to the queue and is taken again on the next pass, but no more than five attempts in total, after which the field stays unchanged and the run continues; in a single-item run the field stays unchanged. Then run again later or choose another model
    “Incorrect API key provided: sk-proj-…” Your own key was entered with a mistake or has been revoked Fix the key in the API settings
    “Project … does not have access to model …” Your own key is valid, but the project in the service's account doesn't have access to the selected model Choose another model from the list or grant access to this one in the service's account
    You ran it for products, but blog article titles also show up in the log The launch window opened from a template shows all four blocks, and in the “Articles” block the same four fields are checked by default Uncheck the fields in the “Articles” block before the run — model requests for them are paid too (starting a run)
    The text came out in a different language than expected The language is set by the main catalog language, not by the launch window Change the main language in the “Languages” directory and run again (about language)

    FAQ

    Why does a run take hours?

    Each field of each item is a separate request, and the queue takes 10–40 items per pass depending on the field. There's no cap on the number of items, and Elbuz doesn't estimate the time in advance (run limits).

    Some products were left without a description — why?

    After the fifth failed attempt an item is no longer taken, and with “Only update empty fields” the answer doesn't overwrite a field filled in by hand during the run. Run it again: with “Only update empty fields” it takes only the empty ones.

    What happens if I click “Run” twice?

    A repeated run of the same field for the same scope replaces the unfinished previous job: unprocessed items aren't duplicated.

    Can I generate with my own API key?

    Yes, add the key with the “Add API access key” button. The cost line under it always shows €5.67 per thousand descriptions — go by your service's price list (cost).

    Can I close the window while generation is running?

    Yes. The run works in the background through the job queue, and closing the window doesn't affect it. You can follow progress on the “Log” tab or in the “Job Queue” window.

    Where to find it and who has access

    Menu
    text generator → template → “Start generation”
    Other entry points
    category card → “AI generation”; base catalog → “AI & content” or the row menu
    Access right
    access to the “Text generator” window
    Prepare
    tested prompts and money on the balance

    All entry points are listed in the article about the generator.

    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