Tripling Product Search Relevance Across Every Market with Roundforest
X2.2
better multilingual support
10
localized versions across countries
300%
list relevancy improvement
Overview:
- An e-commerce platform runs an AI-powered shopping tool that lets consumers compare products, get advice, and find deals across multiple markets and languages.
- Its existing model translated search terms literally, so it missed the phrases shoppers actually use, misapplied filters once translated, and couldn't surface relevant related products. Search that worked in one market degraded in the next, capping personalization and upsell.
- We rebuilt the engine using advanced generative AI and multilingual techniques. List relevancy tripled, translation accuracy climbed from 35% to 76%, and personalized recommendations now run at scale across every market RoundForest sells in. The result is search that converts in any language and a foundation for further AI capability.
GenAI that could read search intent across languages, apply translated filters, and surface related products
Together with Latent, we built an engine that reads a shopper's intent and surfaces what they genuinely want, and the results speak for themselves: tripled list relevance. They were precise and focused on the outcome, and left us a foundation we keep building on."
RoundForest sought to solve three key areas that their existing translator couldn't solve. The existing translators lacked contextual and semantic understanding, and their literal translations missed popular local search terms. When multiple translated search filters were applied, the application was incorrect and the filters were not able to generate relevant related products from a user's search.
GPT based system using advanced generative techniques. Plus, multilingual prompting to structure outputs and personalize at scale
We rebuilt a search engine on a modern LLM stack that structures outputs and user personalization. Drawing on advanced prompting and reasoning techniques, we got the model to return structured, reliable outputs, apply filters correctly, and break categories down meaningfully, while multilingual techniques extended the same quality across languages.
Tripled product search relevance across languages
Latent tripled Roundforest's search relevance and made it hold in every market it serves. By translating product search into ten localized versions across Roundforest's target countries, then layering filtered search with constrained upsell and cross-sell, the system now generates relevant related products straight from a user's search results, in any country Roundforest operates. Translation accuracy rose from 35% to 76%, and list relevancy improved 300%.
