Suppliers and third-party sellers deliver product data in their own formats, and customers were not finding the right products. An ML model on text embeddings places each product in the right category, and small language models turn the supplier or seller text into a product detail page with structured technical attributes.
An e-commerce platform sells products from two sources: suppliers delivering its own range, and third-party sellers listing on the platform. Both deliver product data in their own format: their own category names, their own columns, and descriptions written for a price list rather than for a customer comparing products on screen.
Before a customer can find a product, two things have to happen. The product has to sit in the right place in the platform's category tree, and the supplier or seller text has to become a product detail page in the platform's own format, with the technical attributes customers look for.
Customers were not finding the right products. Products placed in the wrong category, or in a broad catch-all one, never appeared where customers were browsing. Product pages carried the supplier's or seller's wording, and technical attributes stayed buried in free text, so a customer narrowing a search by specification did not see those products at all. Customers who could not find what they came for dropped off, and the revenue went with them.
Every new supplier and seller added another format to map. The products were in the catalogue but could not be found: the problem sat in the product data, upstream of search, because search and category navigation can only surface what the data describes.
Two stages, run on every product a supplier or seller delivers, then a production rollout.
We generate text embeddings from the supplier and seller catalogue data, and an ML model trained on those embeddings assigns every product to its category in the platform's tree. Embeddings capture what a product is rather than the words a supplier or seller happens to use, so two sources describing the same item in different terms land it in the same category.
Small language models rewrite each supplier or seller description into the platform's product detail page format, so every page follows the same structure regardless of where the product came from. In the same step they extract the technical attributes from the free text into structured fields the storefront can filter and compare on.
The task is narrow and repeats for every product in every catalogue, which is the kind of work small models tuned to one output format do well.
Both stages went into production on the live platform, and we measured the effect on how customers engaged with the products they were shown.
The uplift came from the product data. Customers found the right products because those products sat in the right category and carried the attributes customers search and filter by.
Let's talk about what's achievable for your specific context.
Start the Conversation →