One run pairs the price list rows your rules missed. The report gives each pair's confidence and how the products are alike and differ. Apply the pairs; add the rest as new products.
When you need this:
- Leftovers after an import — the model compares the model code in the name, part number, color, capacity, size.
- A regular supplier — automatic mode applies pairs above the threshold after each import.
- New products — “No match” rows with a category go to the catalog with one button.
- Many price lists — one combined run and report.
Where to start: source → “Product Matching” arrow → “AI product matching”.
What AI matching does
The model only gets the leftovers: rows that the matching rules could not pair during the import. A run produces a report, not ready-made links: nothing changes in the catalog or the price list until you click “Apply checked”.
- Pair suggestions with an explanation — Elbuz picks candidates, the model chooses one and writes how the products are alike and how they differ.
- A report with match types — “Exact”, “Similar”, “Suggested”, “No match”, with filters and word highlighting.
- One-click apply — pairs above the threshold are already checked.
- New products to the catalog — “No match” rows with a category are added as new products.
- Automatic mode — a run and apply after every price list import, with no operator involved; the confidence threshold and other settings are described there too.
- Combined run — all price lists, one report.
- Assistant in manual matching — an assessment of the selected candidate right in the manual matching window.
How to start a run
The control panel on the left has a “Run AI matching” button, and the number in brackets next to it shows how many price list rows will be processed. Elbuz asks “Run AI product matching?”, and once you confirm, the job is queued.
1
Run AI matching2
Keywords field3
Settings4
Manual mode- Run AI matching — the number in brackets is how many price rows will be processed.
- Keywords field — selects products for a run and, after a run, searches the results table.
- Settings — threshold, number of candidates, automatic mode.
- Manual mode — switches to manual matching.
A run includes price list rows that meet all three conditions:
- they don't have a match in the catalog yet;
- their price is greater than zero — there's no point matching a product without a price;
- they haven't been through the model in previous runs.
Rows that are already linked — correctly or not — are not reviewed by the model.
The third condition explains the most common confusion with this window.
The counter next to the button shows zero, and trying to start gives “No products to match (product_id = 0)”, even though the price list clearly has unmatched rows. The reason: those rows are already in the report from the previous run. They aren't processed again, so you don't pay for the same work twice.
To rerun them, click “Clear” — the ✕ icon in the action row above the table, right of the “Matched by filter” counter. This clears the results log: unreviewed rows and the history of applied records go, while the products themselves and the links already created stay intact. After that, the counter shows the number of rows again.
To process only part of the price list, use the “Keywords to filter products (optional)” field. It searches by name, manufacturer and category; separate several words with spaces, and words shorter than two characters are ignored.
An empty field means “process all products”. The same field also searches the loaded results table, so after a run you can simply use it to navigate the report.
1
Keywords field- Keywords field — selects products for a run and, after a run, searches the results table.
During a run you see a progress bar and the line “Processed: N / M (X %)”; if some rows failed, “Errors: K” appears next to it. The “Operation log” block expands by itself and shows the model's latest decisions as lines like “supplier name → catalog name [85 % — Similar]” or “supplier name → No match”.
You can close the window: the run continues in the background. When you open the window again, it picks up the job and the log shows “Resumed running job”. Elbuz won't start a second run for the same price list until the first one finishes: it says “AI matching is already running for this price list”.
Next to the run button is “Manual mode”: it opens the manual matching window for the same price list, for when it's clear that you'll sort out the leftovers faster yourself.
To stop a run, click “Cancel job” — it's visible only while a run is in progress. The dialog says “Stop the running AI processing? Already processed results will be kept.”, and the “Also remove unreviewed results (so you can restart immediately)” checkbox removes what has accumulated along with the job. Rows you've confirmed are not affected.
Model requests are paid; usage is charged to the account balance.
How the model decides that products are the same
Before calling the model, Elbuz picks candidates: for each row it runs a smart catalog search by name and manufacturer. It starts within the brand: Elbuz finds the manufacturer in the catalog's manufacturer directory and searches only that manufacturer's products.
The whole catalog is searched when the manufacturer isn't in the catalog, the brand yields fewer candidates than configured, or the products found share no name words with the row — say, the brand is spelled differently in the price list and the catalog. Category and part number don't narrow it; the model filters.
There are ten candidates by default, twenty at most. No candidates — no model request: the row gets “No match” right away. If the right product isn't among the candidates, it's also “No match”, even when the product is in the catalog — so the more the matching rules link, the smaller the leftover for the model.
The model receives text only: from the supplier side — the name, manufacturer part number and model; for each candidate, also the internal catalog SKU.
The model deliberately doesn't get the manufacturer: it takes the brand from the name, since price list brand columns are sometimes wrong. It sees no prices, stock, descriptions or photos. One request covers one price list product with its own list of candidates.
The model's rules are set by Elbuz, not by settings:
- the model code in the name decides; case, spaces and hyphens are ignored, so “WMD 44MB” and “wmd-44mb” are the same. A full model code match gives a high-confidence pair even when the part numbers differ;
- a matching part number only adds confirmation, and different part numbers don't cancel a model code match: on one side the part number may be an internal code;
- similar names with different models are different products; H540 and H390 are not the same;
- any differing variant attribute makes it a different product: color, memory or capacity, size, package contents or supply region;
- if there are several candidates and no exact model code match, the model must answer “no match” instead of guessing;
- the model's own confidence below 60 % also gives “no match”.
A separate check guards against made-up answers. If the model names a product not on the candidate list, the decision is rejected with the reasoning “AI returned a product outside the candidate list (ID …) — decision rejected” with the suggested product's ID. Such a row gets no pair.
The model writes its comments in the interface language: an English interface gives English explanations.
- Action or result
- Decision: the path depends on the answer
Prices play no part in the decision at all. The model treats two products whose prices differ threefold as one if the model code in the name matches and the variant attributes don't conflict. An identical price, in turn, won't make a pair out of two different products.
This matters when similarly named catalog items differ only in package contents or supply region. The model tells them apart by name and part number — if the supplier left that difference out of the name, nothing tells them apart. Check such pairs yourself.
The model doesn't see candidates in another condition. For a discounted, used or refurbished row, the model gets only catalog cards in the same condition: it would treat the word “УЦІНКА” (discounted) in the name as extra noise. With no such cards, the row gets “No match” and can become a new product. Details: Product condition.
How to read the report
At the top of the window are six tiles: “Total”, “Exact”, “Similar”, “Suggested”, “No match” and “Confirmed”. The tiles are clickable: click “Similar” and the table shows only those.
Below the tiles is a summary row with five metrics: “Catalog coverage”, “Average confidence”, “Confirmed”, “AI calls” and “Last run”.
Two of them are easy to misread. “Catalog coverage” is the share of price list rows the model found a pair for; it's calculated from the current report, so it drops after you apply pairs. “AI calls” is the number of rows that had candidates: it shows the amount of work, not spending.
Between the filters and the table is the action row, home to almost everything this article mentions: the “Matched by filter” counter, two icons “Check all” and “Uncheck all”, the “Apply checked” and “Add new products to catalog” buttons, then three unlabeled icons — refresh the table, show the “Operation log”, and ✕ “Clear”. Hover to see each label.
1
Stat tiles2
Metrics row3
Checkbox column4
Score column5
Type column6
Table action row- Stat tiles — a click sets a filter by match type.
- Metrics row — catalog coverage, average confidence, confirmed, calls, last run.
- Checkbox column — confirmed or rejected; items above the threshold are pre-checked.
- Score column — the model's own score.
- Type column — exact, similar, suggested, no match.
- Action row — Apply checked, refresh, operation log and ✕ Clear icons.
The table has five columns you read: the status checkbox, “Supplier product”, “Catalog candidate”, “Score” and “Type”; the rightmost column selects rows for bulk actions. Cells show photo thumbnails of both products. Sort by clicking a column header; by default the most confident pairs are at the top.
The percentage in the “Score” column is the model's own estimate; Elbuz doesn't do any math on top of it. The match type maps directly to the percentage:
| Type | Score | What it means |
|---|---|---|
| Exact | 90–100 % | The model code in the name fully matches, the variant attributes match; the part number matches or is missing. |
| Similar | 70–89 % | The model code matches, but the supplier's name is incomplete or has extra words. Worth a look. |
| Suggested | 50–69 % | The name is similar, but the model code in it is hard to read with certainty. Always check. |
| No match | 0 % | There's no matching product in the catalog. You can add such a row as a new product — set a category first. |
“No match” rows always show zero. The model may have been very sure there's no pair, but showing that certainty as a percentage would confuse it with confidence in a pair.
The model's comment has no column — it's in a tooltip. Hover over the catalog product name to see why it's the same product. For rows without a pair, it's behind the “?” icon next to the supplier name. Hovering over a row also highlights words: green for what matches in both names, amber for differences.
Above the table are five filters: “Match type”, “Confidence, %” with a “from” — “to” range, “Price row” (any, without category, with category, deleted from the price list), “Review status” and — in combined mode — “Price list / supplier”. Next to them is “Reset filters”. A page shows one hundred rows.
Two things in the table look like data loss but aren't.
- applied pairs disappear from the report completely: the link record lives on in the “Matched products list”, not here. Under the “Review status” → “Applied” filter only checked “No match” rows remain — the ones that had nothing to link to.
- Dimmed rows with the tooltip “Price list row no longer exists” are records whose row disappeared when the price list was reimported; all you can do with them is delete the record.
How to review and apply pairs
The checkbox in the first column is your decision on the pair. Checked — the pair is confirmed; unchecked — rejected. The table doesn't reload, so you can move through the report quickly.
Not only “Exact” pairs are checked automatically. Everything that reaches the confidence threshold gets a checkmark — 60 % by default, which includes “Similar” and “Suggested”. “Exact” pairs are always checked, regardless of the threshold.
So clicking “Apply checked” right after a run also applies pairs that Elbuz marked as worth a look. A wrong pair links the row to the wrong catalog product, and the same mistake immediately spreads to identical products in other price lists. Go through the checked rows: that's what the model's comment under each pair is for.
The two icon buttons “Check all” and “Uncheck all” work strictly on the current filter, not on the whole price list. A handy approach: filter by “Exact” and check everything, then switch to “Similar” and go through them one by one.
The “Apply checked” button with a number in brackets does the main work. Elbuz asks “Apply all checked matches to price list products?”, and then the price list rows are linked to catalog products — by the same mechanism as manual matching.
The same product in other price lists gets linked at the same time. At the end you see the summary “Applied (linked / to be created)”: the first number is links created, the second is checked “No match” rows that have nothing to link to.
When applying, Elbuz checks the product condition once more — this concerns suggestions saved before the condition rules changed. If the row and the suggested card are in different conditions, no link is created: the row becomes “No match” with the reason “Different condition (new / discounted / used) — the product was not linked.”
How to undo a decision before and after applying
You can undo a decision at two different stages, and the cost differs.
Before you click “Apply checked”, just uncheck the box. The pair simply isn't created: the record becomes “Rejected”, and nothing changes in the catalog or the price list. There's also the “Reset the picked product” action — it returns the row to “No match” and clears the candidate.
After you click it, the link already exists and sits in a separate links directory. You can't see or remove it from the matching window: the record moves to history. Break it in the “Matched products list”: arrow next to the “Product Matching” button in the source header → “Matched products” → trash icon in the row.
Unchecking now is cheaper than hunting for the link later. A doubtful pair in the report is one click; the same pair after applying means switching to another window, finding the row among all the price list's links and deleting it by hand.
How to add new products to the catalog
A “No match” row becomes a new product through its category, not a checkmark: you don't need to check it, just don't reject it and set a category. The “Add new products to catalog” button appears when the report has such rows and adds them to the base catalog as new products.
- There's one condition, and it's strict: the row must have a category. The dialog warns you about it before starting: “Only products with an assigned category will be added. Products without a category must be categorized first.”
- When it finishes, you see “New products added to catalog: N”, and the operation log shows “New products added to catalog: N / skipped without category: M”.
- The window has no separate list of skipped rows: they stay in the report with the “No match” status. The “No match” tile + the “Price row” → “Without category” filter show exactly those.
- Set the category one by one or for a group of selected rows at once with the “Set category” action; AI can also pick it if “After matching, assign categories to new products” is on.
- Identical new products from several suppliers in the combined report become one card — the log shows this as “duplicates merged: K”.
From the run to new products in the catalog
What you can do with a single row
Action icons appear when you hover over a row. There are four above the supplier name.
1
Pencil2
Magnifier3
Ban sign4
Cross5
Return arrow- Pencil — edits the supplier product name in the row; double click does the same.
- Magnifier — searches the name in Google.
- Ban sign — an ignore rule that deletes rows with this name from the price list.
- Cross — removes the entry from the report, price list and catalog stay untouched.
- Return arrow — resets the chosen product and returns the row to No match.
- “Edit supplier product name” — edit right in the row; double-clicking the name does the same. The new name is saved in the price list; Elbuz won't accept an empty name. Useful when the supplier sent a name with junk that keeps the candidate search from finding anything.
- “Search product name on Google” — opens a search in a new tab when the name doesn't make clear what the product even is.
- “Add product to ignore list (remove from price)” — a stronger action than it looks: it creates an ignore rule by name and deletes rows with that name from the price list. For junk in the price list that's exactly right, but you'll have to undo it in the ignore rules.
- “Delete” — removes the record from the report without touching the price list or the catalog.
Above the “Catalog candidate” column, the set depends on whether there's a pair. If there is, “Reset the picked product (back to “No match”)” is available. If there isn't, there are two other items.
“Pick a catalog product manually” opens the “Pick a product from the base catalog” window: the search field is already filled with the supplier name, search works from two characters, and clicking a found row links the product. The product you pick is saved as a 100 % “Exact” pair, marked as chosen by the operator.
“Set category to add to the catalog as a new product” — the same step without which the row won't get into the catalog as a new product.
When there are many rows, it's easier to select them with the checkboxes in the right column — a bar with three bulk actions then appears under the table: “Set category” for all selected rows, “Ignore” and “Delete”. The first one is the most useful: categories for new products are usually set in batches.
How to process all price lists in one run
You don't always need to open each supplier separately. The same window can work across all price lists at once and produce one combined report — this is an everyday workflow, not a rare feature.
In “List of price lists”, switch to “All sources”, open the “All products” tab and choose “AI product matching” in the product panel menu. The globe icon marks combined mode, but both AI items — products and categories — have it, so go by the name. With a source selected, the item opens that source only.
1
All sources2
AI product matching3
Globe icon- All sources — without this mode the item opens the window for one source.
- AI product matching — a combined report for all price lists at once.
- Globe icon — distinguishes the combined mode from the single-source mode.
The window is the same, but behaves differently in a few places.
| What | One price list | All price lists |
|---|---|---|
| Counter next to the run button | number of rows to process | zero: counting across all price lists is expensive and not needed just for viewing |
| Settings in the “Settings” window | this price list's settings | general settings |
| What a run creates | one job | a separate job for each active price list that has something to process |
| Progress line | “Processed: N / M (X %)” | the same plus the number of jobs |
| “Price list / Supplier” column | none | present, collapsed; a toggle in the panel shows it, and the choice is remembered |
| Price list filter | none | present, with a list of active price lists |
| “Apply checked” and “Add new products” | for this price list | for all price lists at once |
In combined mode, progress also counts jobs that finished in the last half hour. Otherwise, when the first job closed, the counter would roll back and it would look as if processing had started over.
In combined mode, the “Clear” button clears the results log for all price lists, not just the one you're filtering by.
The same goes for “Cancel job”: it stops all active runs. And for “Apply checked”: it links all checked pairs in all price lists, even if the filter currently shows one price list; if nothing is checked at all, Elbuz replies “No confirmed records to apply”. To remove one supplier's results, open the window from that supplier's source.
How AI helps in the manual matching window
The model also helps when you sort out the leftovers by hand. In the manual matching window, the menu has the “AI assistant (one AI request per product)” item — on by default. The assistant looks at the price list row you've stopped on and tells you what it thinks of the selected candidate.
The request goes 1.8 seconds after you stop on a row, so fast scrolling doesn't queue requests. Returning to the same product doesn't repeat the request: the answer comes from the window's memory or, if the row has been through a run, from the report. The assistant doesn't work in the window's compact mode: there's nowhere to show the answer.
There are three kinds of answers:
- The model agrees with your choice — the candidate card shows a chip with a robot and a percentage, and under the name there's a comment on why it's the same product.
- The model suggests a different product — a collapsed line “AI suggests a different product” appears with a percentage and a “show” link. The expanded card shows the photo, name, comment and a “Show” button that highlights the suggested product in the candidates grid. If it's not in the current list, Elbuz tells you.
- No pair — a collapsed line “AI found no suitable product” that expands into a comment.
You'll also see the yellow bar “AI matching is running for this price list” with progress and the warning “Manual changes may conflict with AI results right now.”. The bar shows while a background run is going on for the price list and disappears when it finishes. The “Open AI window” button on the right takes you to the report.
Scenarios
The field labels below come from the “AI product match settings” window, opened by the “Settings” button in the matching window panel. What you set there applies to the current price list.
New products from a supplier should reach the catalog with a category already set
What to do. Turn on “After matching, assign categories to new products”. When the report is ready, click “Add new products to catalog”.
What you get. After a run that finished without errors, AI category matching starts for “No match” rows without a category: first by the supplier's category names, then by product names.
Categories with confidence above the category AI threshold are set for new products automatically. If the same product — same manufacturer and same name — came as a new product from several suppliers, the button in the combined all-price-lists window creates one card in the catalog and links the other rows to it.
What won't happen. Category matching doesn't add new products to the catalog — it only prepares the rows for the button. Rows that still have no category are skipped by the button and counted separately: set their category by hand, one by one or with “Set category” for a group of selected rows.
Process only one brand or one product group first
What to do. Before starting, enter the brand name or a word from the supplier's product name or category in the “Keywords to filter products (optional)” field and click “Run AI matching”.
What you get. The run includes unmatched rows where at least one of the words is found in the supplier's name, manufacturer or category. Several space-separated words widen the selection rather than narrow it: a row matches if any of them is found. If nothing matches the filter, Elbuz says “No products match the keyword filter”.
What won't happen. The rest of the price list isn't processed or marked: the next run with an empty field takes the rows that aren't in the report yet, and the ones already processed aren't run again.
FAQ
Why does the counter next to the run button show zero?
Those rows are already in the report from the previous run and aren't processed again. Click “Clear” (✕ above the table): the results log is cleared, while links and products stay. “No match” rows return to the run by themselves after the set number of days (starting a run).
Are the automatically checked rows only “Exact”?
No. Everything that reached the threshold is checked — 60 % by default, so “Similar” and “Suggested” too. Review them before “Apply checked” (reviewing pairs).
I applied a pair by mistake. How do I remove it?
Not from the AI window anymore: an applied link lives in the directory. Open “Matched products” from the “Product Matching” button menu and delete the link with the trash icon (breaking a link).
Does the model take the product price into account?
No, prices aren't passed to the model. The decision is based on the model code in the name, the manufacturer part number and the model; the price only filters out rows with a zero price (how the model decides).
Why won't the run start, with a message saying AI is unavailable?
“The daily demo limit is used up…”: continue tomorrow or on a paid plan (demo: €3/day). “No AI access: AI service API key is not configured.” means the account has no service key set up. In that case no job is created and no price list rows are used up.
Elbuz says “No confirmed records to apply”. Why?
The report has no checked row that hasn't been applied yet: applied rows don't count a second time. Check the rows you need or use “Check all” for the filter; if you can't see any checkmarks, check the report filter (reviewing pairs).
I clicked “Add new products to catalog”, but some products weren't added. Why?
The button adds only “No match” rows with a category; the rest are skipped and counted in the operation log. Another reason: identical new products from several suppliers are merged into one card (new products to the catalog).
Can I confirm all pairs in one click?
Yes, within a filter: set a filter, for example “Exact”, click “Check all”, then “Apply checked”. There's no separate “Confirm all” button: confirmation is the checkbox in the first column.
Where to find it and who has access
- Other entry points
- “List of price lists” → source → arrow next to “Product Matching” → “AI product matching”
- From the source card
- the “Matching” tab — the same window, embedded in the card
- All price lists
- “All sources” → “All products” tab → menu → “AI product matching”
- During a run
- yellow bar in the manual matching window → “Open AI window”
- Access right
- the “List of price lists” window; the “Matching” tab — the price list processing rules block
The “Product Matching” button in the source header has two parts: the left part opens manual matching, the arrow on the right opens a menu with five items. “AI product matching” is the first one.
| Menu item | What it opens |
|---|---|
| “AI product matching” | the AI matching report; opens maximized |
| “AI category matching” | model-based matching of price list categories — a separate topic not covered here |
| “Attribute mapping” | checks price list attributes against catalog attributes |
| “Matched products” | “Matched products list” — all links created so far |
| “Price list category matching” | manual matching of the price list category tree |
1
Left half of the button2
Caret3
AI product matching4
Matched products- Left half of the button — opens the manual matching, as before.
- The caret hides the other matching windows.
- AI product matching — first item, opens the full-screen report window.
- Matched products — the directory of created links, where you break a link after applying.
Related topics
- If the leftover for the model is too large — set up automatic matching rules, they run earlier.
- If there are only a few pairs and you'd rather decide yourself — How to manually match price list products to your catalog.
- If new products need categories — AI category matching.
- If an applied pair needs to be broken — Matched products list.
- If you're deciding whether to hand price lists over to Elbuz — what it gives your store.
