A model for ranking catalog search results

You assess the whole path to the product: whether the right product is on the list, in what position and whether the customer reaches it. A model that sets the order will not help if the search does not find the product at all, because the description of the use matches no word on the product sheet. Then the first step is better names, synonyms and descriptions of uses.

First find, then set the order of catalogue search results

Say a customer searches for “a desk lamp with a clamp”. The catalogue has a suitable product, but the search shows ordinary standing lamps first, and the clamp model only further down. If the right item is already among the candidates, you can assess a change to their order. If it is not, you first need to check the product description, the filters and the search itself.

This difference affects the scope of the order. A model that orders the list it receives will not lift a product it was never given. Improving synonyms, or adding a missing feature in the catalogue, can solve the problem earlier, with no extra training.

Sentence Transformers describes Cross Encoder models used to re-rank candidates found by a search engine. In this use, the model scores a pair: the query and the description of a result. The team can compare that order with the current results.

What the team counts as a relevant result

The catalogue owner should explain why the lamp with a clamp answers the query, while an ordinary standing lamp lacks the feature asked for. Keep the queries where several products fit as well. A relevance assessment does not have to name one item as the only correct answer.

A product's popularity and promotion are separate commercial decisions. If the company changes the order for those reasons, that has to be kept apart from the assessment of fit with the query. Otherwise an employee will be correcting the model by a criterion nobody agreed beforehand.

A ready re-ranking model may be enough. Fine-tuning, further training it on checked examples, is worth considering for recurring mistakes about customers' language and the features of the range. The team then supplies the relevance assessments, and the supplier checks whether the adaptation also changes the lists for other queries.

A purchase assessed on the whole path to the product

The comparison should cover queries where the current search fails and those it handles well. The manager sees whether fixing the hard cases makes simple searches by name worse. The wait for the list matters too: a more relevant result has to appear in a time that is useful for the customer.

The first scope can be limited to a chosen part of the catalogue and to queries about uses. You do not need to change the whole shop at once. The client brings knowledge of the products and the decisions on what relevance should mean; Syntalith compares the available approaches and prepares the result for the team's assessment.

Describe a query where the right product gets lost in the results. We can start the conversation about search from it. Information about working together is on Syntalith's pricing page.

What a search ranking 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.