Elbuz pairs up similar cards among the selected products, scores each pair from 0 to 100, and lets you mark it: duplicate or not. Exact matches confirm themselves.
When you need this:
- Hundreds of pairs, no time to check by eye — tiles and filters narrow them by score, key, or name.
- Names nearly match, products differ — in “With AI check” mode the model sets variants aside.
- A pair was rejected last week — later runs hide it.
- The AI check stopped halfway — a new run checks only unchecked pairs, with no double billing.
Where to start: open the window from the base catalog and click “Find duplicates”.
What the pair search can do
- Checks the grid's filtered set — the run takes what the base catalog shows right now: category, filters, checkboxes.
- Two modes — “Without AI” is free, “With AI check” adds the model's verdict for every pair.
- One button for the whole cycle — “Find duplicates” compares products and passes the pairs straight to the AI check.
- Pair review — buttons in each row, bulk actions with checkboxes, pages of a hundred pairs.
- Auto-confirmation — exact matches and a confident AI verdict become “Confirmed” on their own, without deleting any cards.
Which products get checked
The window takes the grid's at the moment it opens: the tree category with its subcategories, the filters, the search box, and the checkboxes. The first number in the header, “N products selected”, is how many go into the run, unless the run limit cuts them off (see below).
Next to it, the “Condition” field reads “ticked in the catalog” if the grid had ticked products, and “by catalog filter” otherwise. How to narrow the filtered set is described in the catalog grid article.
If nothing is filtered at all, the window warns: “No scope set — the whole catalog will be checked. To narrow it down, set a filter or tick products in the base catalog.” This is not an error: a full run is needed and useful, it just takes longer.
Run the first search on the whole catalog, even if you plan to work category by category later. A pair makes the list if at least one of its cards is in the filtered set, but the other card must already be processed. After a full run, checks of single categories also catch duplicates that sit elsewhere.
A single product is recalculated automatically: as soon as you save its card, its name, part number, and model keys are rebuilt. You don't need a full run for one edit.
What never goes into a run: variant products. Cards linked to a main product are not treated as duplicates.
One run is capped at 200,000 products. If there are more, you'll see: “Checked N products out of TOTAL — the rest did not fit into a single run. Narrow the scope with a catalog filter to check the remainder.” The cut is reported, but the remainder must be run separately. In a demo account the cap is 100 products.
Which mode to choose: without AI or with AI check
The window header has two toggles: “Without AI” and “With AI check”. They differ not only in speed and cost: each groups the list of found pairs differently, because each has different data to group by.
- You need it fast and freeThe list is grouped by name similarity, the operator decides
- You need accuracy, and the catalog has many variantsAI looks at every pair and sets variants of one product aside
Without AI
Elbuz shows similar pairs, and you decide on each one. It costs nothing. There are six tiles above the table: “Very similar” (score at or above the duplicate threshold), “Similar” (between the two thresholds), “Confirmed”, “Merged”, “Rejected”, and “Pairs total”.
With AI check
The model looks at every pair found and answers whether it's one product or two. It's slower and costs money, but the list is grouped by verdict: “Duplicates (AI)”, “Awaiting AI check”, “Variants”, “Different products”. Pairs go in batches of 20 per request, and the window repeats the step until the list is done.
There are exactly three verdicts: duplicate (the same product), different variant (the same model line, but a different volume, color, size, or package contents), and different products. When in doubt, the model answers “different products”: a false “duplicate” costs more than a missed one, because the operator would merge two live cards based on it.
What you pay for. The billing unit is one AI request, and a request holds up to 20 pairs. The request price is based on tokens: what was sent to the model and its reply are billed separately, at the model's rates.
For each card the request includes the name (up to 200 characters), plus the manufacturer part number and brand if filled in, so long names cost more. The check runs on the GPT-6-Luna model. If your settings have your own OpenAI key, the request goes through it, to the model set for that key.
You pay only for pairs the AI hasn't seen yet: a pair with a verdict is left out of later requests, and a new run doesn't erase its verdict. If the model skips a pair in its reply, the pair stays without a verdict and goes into the next batch, that is, the next paid request.
The window itself shows no amounts. Each batch is logged as a separate row in the Operation log with a “Cost” column; the row is labeled “Product duplicate check (AI)”. Log rows are kept for 14 days.
Why bother if there's already a score: the score measures name similarity, and that can't tell a variant from a duplicate. Among pairs scoring 85 and above, 18 turned out to be duplicates and 74 variants.
Example: “QUBE QBC-GPM-650W-80B” and “QUBE QBC-GPM-650W-80BBULK” have nearly identical names but are different products; only someone who knows what the suffix means can tell. With default settings this pair won't even reach the list: its score is 46, below the doubtful zone threshold.
In a demo account the AI mode works within a daily limit; once it is used up, the window shows “The daily demo limit is used up: up to €3 a day for AI, parsing and media generation. You can continue tomorrow or on a paid plan.”
1
Duplicates (AI)2
Awaiting AI check3
Variants- Duplicates (AI): pairs the AI confirmed as one product, ready to merge.
- Awaiting AI check: pairs found but not yet sent to the AI.
- Variants: pairs the AI judged to be variants of one product.
How to start and stop the search
One button covers the whole cycle. In AI mode it first compares products, then passes the pairs found straight to the check, and turns into “Stop” while the check runs. A batch already sent to the server finishes and saves its verdicts; the stop takes effect before the next one.
After a stop: “Check stopped. Pairs checked: N. The remaining doubtful pairs stay in the list — you can continue later.”
If a batch request fails (the server didn't respond in time or returned an error), the check also ends: the window shows “Pairs checked: N” and the error text in a separate red message. Pairs that weren't reached stay without a verdict, and the next run will check them.
- Action or result
- Decision: the path depends on the answer
A stopped check is resumed with the same button. Clicking “Find duplicates” again starts a new run: it removes pairs that are awaiting review and have no AI verdict yet, searches for pairs anew, and sends only the pairs the AI hasn't seen to the check.
Pairs that got a verdict before the stop stay in the list with their verdict, and you don't pay for them again. Your own decisions (confirmations, rejections, merges) aren't touched by the run.
Below the header, the window suggests the next step, and the line changes with what's already done. Before a run: “Press ‘Find duplicates’…”; while it works: “Comparing products…”.
In “With AI check” mode, while the list has pairs the AI hasn't seen, the hint suggests running the search again: “Some pairs are still unchecked — the check was stopped.” In “Without AI” mode this line never appears: the hint suggests reviewing the pairs instead.
As soon as the list has at least one confirmed pair (in AI mode, once no pairs without an AI verdict are left): “There are confirmed pairs — press ‘Apply’ to merge them.”
1
Products selected2
Without AI / With AI check3
Find duplicates4
Search settings5
Very similar6
What to do next- The number of products selected in the Products window; the search runs on exactly these.
- The mode chips: fast Without AI or slower With AI check.
- Find duplicates starts the search over the selected products.
- Search settings opens the thresholds dialog.
- The Very similar tile counts the pairs with the highest score; click it to filter the table.
- The hint line under the tiles says what to do with the current list.
Along with the result, an amber warning may appear: “Some products could not be checked”. It means the same value sits in an identifier field on more than 20 cards, for example a part number of “0” shared by many products. Such a value can't tell one product from another, so the whole group was left out of the search.
Elbuz lists these values with the number of cards: either fix the field in the catalog, or turn the field off in the search settings.
How to review pairs
The tiles above the table are filters. Clicking a tile shows only its pairs. In AI mode “Duplicates (AI)” opens first: it holds the list ready for merging.
Below the tiles is a filter bar, and on a large catalog it's the main tool, not the tiles. “Matched by” keeps pairs found by one specific key; “Score” “from” and “to” narrows the list to the band you need; “Search by name” looks for a substring in both names of a pair.
All five column headers sort the list (by default, “Score” from highest to lowest), and each page shows a hundred pairs: if the list seems to break off, use the pages at the bottom.
Besides checkboxes, the table has five columns. “Score” is a percentage plus what found the pair: “EAN barcode”, “manufacturer part number”, “name”, “a rare word of the name”. “Card 1” and “Card 2”: thumbnail, name, “ID”, part number. The comes first for a reason: names often match word for word, and only the ID tells cards apart.
“AI check” shows the verdict, a confidence percentage, and an icon that hides the model's explanation; in “Without AI” mode it reads “not checked”. The last column is “Status”.
When you hover over a row, matches in the names are highlighted: green for what's the same, yellow for what differs. Highlighting appears only on the hovered row: in a dense table, permanent shading turns into noise.
On hover, “Status” also shows three buttons: “Confirm”, “Merge”, and “Not a duplicate”. A confirmed pair gets “Clear decision” instead of the first one: it returns the pair to review and out of the “Apply” queue. The card cell has an “Open card” button: it opens the product card in a separate window to compare both cards before deciding.
1
Score and match reason2
Highlighted words3
Card ID4
Row actions5
Row checkbox- The score in percent and, under it, the rule that found the pair.
- Under the cursor, words shared by both names and words that differ are highlighted.
- The product ID and manufacturer SKU under each card name.
- Confirm, Merge and Not a duplicate; the buttons were only shown, not pressed.
- The checkbox in the right column selects the pair.
If there are many pairs, mark them with checkboxes instead of the buttons in each row: tick a few rows, and a “Selected: N” bar appears below the table with “Confirm”, “Not a duplicate”, and “Clear decision” buttons. Merged pairs have no checkbox: there's no decision to clear.
A rejected pair doesn't come back. Once you mark it “Not a duplicate”, later runs won't show it. This saves you from reviewing the same pair every week.
Which pairs are confirmed automatically
Some pairs are confirmed on their own. This happens in two cases: the names produced the same set of words in the same order (a score of exactly 100), or AI answered “duplicate” with a confidence of 95% or more.
A pair gets 100 even if its names are written differently: case, punctuation, a comma or period in a decimal number, and Cyrillic look-alikes of Latin letters are ignored; only words and their order count. Such pairs become “Confirmed” right away, but no cards are deleted: merging is still a separate action. Automation saves marking work without irreversible decisions.
Both thresholds are set in the search settings, in the “Confirm automatically” block: “At score from, %” (100) and “At AI confidence from, %” (95). Zero turns auto-confirmation off, and a value below “Duplicate threshold, %” is raised to that threshold.
Auto-confirmation also cleans up after itself: on a new run, a pair confirmed by automation leaves the list and is recalculated. If the score is 100 again, the pair is confirmed again. If an AI verdict confirmed it, it goes back to review, but keeps its verdict, and AI won't check it again.
The next AI check batch confirms such a pair again: after each batch, Elbuz confirms all pairs awaiting review with a “duplicate” verdict of 95% or more. If there are no new pairs to check, no batch runs, and the pair waits for your decision. Your own decisions (confirmations, rejections, merges) are never touched by a new run.
- Action or result
- Decision: the path depends on the answer
Scenarios
The same product came from two price lists with a different word order in the name
What to do. Open the window from the base catalog with no filters, so the first run covers the whole catalog. In the header, choose “Without AI”. In “Search settings”, keep “Name” and “Rare words of the name” turned on; they are on by default. Click “Find duplicates”.
What you get. Pairs with the same words in the same order score 100 and become “Confirmed” on their own. A pair with the same words in a different order scores no more than 95: with the default thresholds it lands in the “Very similar” tile, and you confirm it. Then “Apply” merges everything confirmed.
What won't happen. No card is deleted until you click “Apply”: automation only moves the pair to “Confirmed”. A pair whose names have numbers that differ both ways (“128” versus “256”) won't make the list at all.
The catalog has many variants that differ by one word or suffix
What to do. In the header, switch to “With AI check” and click “Find duplicates”. AI must be available to the account: a demo account has a daily limit of €3 for AI, parsing and media generation combined.
What you get. After the comparison, the same button sends pairs without a verdict to AI in batches of up to 20. A “duplicate” verdict with a confidence of 95% or more is confirmed automatically and lands in the “Confirmed” tile. Other verdicts are sorted into the “Duplicates (AI)”, “Variants”, and “Different products” tiles and wait for your decision.
What won't happen. Automation never confirms a “different variant” pair, however confident the verdict. You don't pay again for a pair the AI has already checked: a new run doesn't erase its verdict, and it doesn't go into a new batch.
What to do when the search stops or won't start
| What you see | Why | What to do |
|---|---|---|
| “No scope set — the whole catalog will be checked. To narrow it down, set a filter or tick products in the base catalog.” | The window was opened from the tree root with no filters | That's fine for the first run. To narrow it, close the window, pick a category or filter in the grid, and open it again (what exactly gets checked) |
| “Checked N products out of TOTAL — the rest did not fit into a single run. Narrow the scope with a catalog filter to check the remainder.” | The filtered set has more than 200,000 products; in a demo account, more than 100 | Run it category by category: pick a category in the tree and open the window again (what exactly gets checked) |
| “The daily demo limit is used up: up to €3 a day for AI, parsing and media generation. You can continue tomorrow or on a paid plan.” | The demo account's daily €3 limit for AI, parsing and media generation is used up | Work in “Without AI” mode, come back to the AI check tomorrow, or switch to a paid plan (modes) |
| “No AI access: AI service API key is not configured.” | No AI service key is set for the account | Set the key in the system settings or work in “Without AI” mode |
| “Check stopped. Pairs checked: N. The remaining doubtful pairs stay in the list — you can continue later.” | “Stop” was clicked during the AI check | To check the rest, click “Find duplicates” again: AI will look only at pairs without a verdict, and you don't pay again for checked ones (the run) |
| “No pairs selected” | A bulk action was run with no checkboxes ticked | Tick the pairs in the right column of the table (how to review pairs) |
FAQ
How many products does one run check?
Up to 200,000, or up to 100 in a demo account. If the filtered set is larger, the window tells you how many products were checked and how many are left, and suggests narrowing the filtered set with a catalog filter.
Can AI tell an iPhone 15 case from a 15 Pro case? They're different products
Yes, that's what “With AI check” mode is for: the model gives each pair a verdict of “duplicate”, “different variant”, or “different products”. Cases for different models land in “Variants” or “Different products”. In “Without AI” mode such pairs differ only by similarity score (AI mode).
It says “AI check is turned off in the search settings”. What should I turn on?
The “Check doubtful pairs with AI” checkbox in the search settings. It's the same setting as the mode toggle in the window header (search settings).
If I run the search again, will I pay for the AI check twice?
No. A pair with a verdict is left out of new requests, and a new run doesn't erase the verdict. You pay only for pairs the AI hasn't seen yet (what you pay for).
Why wasn't a pair with the same barcode confirmed automatically?
A barcode match raises the score but doesn't make the pair a duplicate: the field sometimes holds junk. Only a pair scoring 100 or with an AI “duplicate” verdict of 95% or more is confirmed automatically (auto-confirmation).
Where it is and who has access
- From the base catalog
- Directories → Base product catalog → Products → toolbar menu → “Catalog duplicates”
- Access right
- the “Catalog duplicates” window
- Prepare
- the products you need, filtered in the base catalog grid
The window has no menu item. Why it opens only from the catalog is explained in the overview article.
Related topics
- If you need the big picture — How to find and remove product duplicates.
- If the pairs are confirmed and it's time to combine them — How to merge two product cards into one.
- If the search misses obvious pairs — search settings.
- If you need the cost of each AI check batch — Operation log.
