Matching supplier products to your internal catalog

The same product, different packaging

Say a supplier offers a pack of six identical markers. The company catalogue has two items: a single marker and a pack of six markers of the same make and model. The names start the same way, and the employee sees the packaging information only after opening each sheet. The code the supplier assigned does not appear in the internal item master.

A helpful suggestion would show both items at once, with the number of units in the pack and the available identifiers clearly visible. The employee could reject the single marker and then check whether the pack really matches the supplier's offer. Similar names do not end that check.

What an identifier refers to matters too. A GTIN is a number that identifies a trade item. GS1 UK explains that each packaging level should have its own GTIN. So the number of a single product should not serve as confirmation of a multipack's identity. Source: GS1 UK.

If the supplier left the packaging field empty, a list of similar products can still help the search. It gives no grounds, though, to choose the single unit or the pack. You have to check the supplier's document or ask them to fill it in. The system should keep this doubt visible next to the proposed items.

When ordinary search is enough for supplier product matching

If both sides have the same verified identifier for the right trade item, looking it up in the item master can settle the matter without a model. A previously recorded link between the supplier's code and the internal item number works the same way. With the next catalogue delivery, the employee does not have to work out the same link again, as long as the codes still refer to the same items.

So first look at why searches fail. Perhaps the code is stored in a different field, the catalogue has duplicates, or approved matches stay only in individual buyers' spreadsheets. Tidying this information can remove a large part of the manual searching. A model is not a precondition for building a shared list of known links.

Items with no shared identifier are harder. The supplier may use a shortened name while the company records a full description, or give the features in a different order. Then a search that takes the meaning of the text into account can find proposals an employee would not see by typing an exact phrase.

What the model has to find

In this use, the model helps find existing items in the catalogue. Its output is a short list to check. In the marker example, the value of the list is that the employee quickly reaches the right pack and sees how it differs from the single unit.

A model can help find descriptions related in meaning despite different words. The supplier and the company's employee may name the same thing differently, and the search will still show the right item to check.

Google distinguishes using models for information retrieval from using them to assess text similarity. Its documentation also describes tuning models on examples specific to a task. That does not prove that similar descriptions mean an identical product. Google documentation: embedding task types and tuning models.

The buyer does not need to choose a search technique before the first conversation with a supplier. They should, however, settle what help the employee is to get. A useful proposal leads to a specific product sheet and shows the information needed to check it. A bare list of similar names leaves much of the work unchanged.

What justifies your own adaptation

First test a ready model on the searches the team finds hard today. If it finds the right items and employees can pick the right one quickly, further training may not be needed.

Adaptation makes more sense when specific mistakes recur: the model does not recognise the abbreviations suppliers use, or puts single units at the top of the list and the matching multipacks much further down. The company then needs reliable examples of correct matches, assessed by a person who knows the catalogue. Records linked in the past also need reviewing, because an old match may have been a mistake.

Before widening the solution, employees should compare the proposals with ordinary search on further items not used to adapt the model. What matters is whether the right product appears on a short list and how long confirming it takes. Also include products the company does not yet have in the catalogue. Good search help should let you close a case with no match, without pushing you to pick the nearest name.

An employee's approval of a link can then save searching for the same item in the next file. The person responsible for the item master still decides whether to merge records, clear up the data with the supplier or create a new item. Checking changed purchase terms is a later step, covered separately in the article on automating supplier price list checks.

A project with Syntalith for one part of the catalogue

Syntalith builds AI apps and models. For a team that maintains the item master, we propose starting with one product group where employees often search for the same or similar items. We check what tidying identifiers and ordinary search solve, and then compare the model's suggestions with them. A trial of model adaptation should show whether the improvement over these simpler methods justifies rolling it out and maintaining it.

The project's result should reach the place where the employee reviews the supplier's offer, so they can open the proposed sheet and approve the link. For the first call, it is enough to tell us how you search for a product today and where you most often get stuck. You do not need to prepare a specification or send the whole catalogue in advance. We price a private model after a call about the data and the hardware, and an app on it starts from €3,000 net (pricing).

What a supplier product matching model costs and what it saves

Where to start

In the free process review we choose one process and 20 to 100 cases the team has already handled. We compare the simpler change with the adapted model on cases the model has not seen, and within 2 working days you get a written summary with a price. We adapt models as part of private LLMs, and build the app as an AI app.