A model for ranking catalog search results
A customer searches for a product by describing what they need it for, but the first results match only part of the wording. The catalog manager wants a better ranking while retaining control over what actually answers the query. A customized model is one option, but first establish whether the right products reach the candidate list at all.
Syntalith
Syntalith proposes an assessment of search and result ranking. Your team can compare current and proposed lists for real queries and explain which products help the customer. The scope covers finding the cause of weak results, comparing simpler changes with possible adaptation, and agreeing how the chosen approach connects to search.
Retrieve candidates before ranking them
Suppose a customer searches for “desk lamp with a clamp.” The catalog contains a suitable product, but ordinary freestanding lamps appear first and the clamp model much later. If the relevant item is already among the candidates, reranking is worth assessing. If it is absent, inspect the product description, filters, and retrieval first.
That distinction changes the work to commission. A model sorting a supplied list cannot promote a product it never received. Better synonyms or a missing catalog attribute may resolve the problem before any further training.
Sentence Transformers describes Cross Encoder models that rerank candidates found by a retriever. The model scores a query together with a candidate description. The team can compare this ranking with the current results.
Agree what a relevant result means
The catalog owner explains why the clamp lamp answers the query and why a freestanding lamp lacks the requested feature. Retain questions for which several products fit. Relevance does not have to identify a single item as the only acceptable answer.
Popularity and promotions are separate commercial choices. If the company changes ranking for those reasons, distinguish them from matching the query. Otherwise reviewers will correct the model against a criterion that was never agreed.
An existing reranker may be sufficient. Fine-tuning means further training it on reviewed examples. Consider it when recurring mistakes concern customer language and product attributes. Your team supplies relevance judgments; the provider assesses whether adaptation changes results for other queries too.
Evaluate the whole route to the product
Include queries that current search handles well as well as difficult ones. The manager can see whether improvement on harder searches damages straightforward product-name searches. Waiting time matters too: a more relevant list needs to appear while it is still useful to the customer.
An initial scope can focus on one part of the catalog and queries describing intended use. The company does not need to replace the entire store at once. Your team contributes product knowledge and relevance decisions; Syntalith compares approaches and prepares results for the team to inspect.
Describe a query where a suitable product gets lost in the results. That can start a conversation about search. See Syntalith pricing for information about working with us.
Match a model to the task you need it to perform
Describe where your current AI falls short. We will compare model customization options, data requirements and the cost of running the resulting system.
Private LLMs and fine-tuning