I recently published work I did at The Knot in which I’m using semantic search for product recommendations. Here is a link to the full article on LinkedIn: Personalized Vendor Recommendations with Semantic Search on The Knot

At The Knot, we have a marketplace with wedding venues on one side and engaged couples planning their wedding on the other side. Finding the right venue is a key challenge to couples and the goal of my work was to make this search easier by leveraging AI in the form of text embeddings.

Summarized at a high level, I built a recommendation system that uses semantic search. The basis for the recommendations are the results that a couple got in our Style Quiz (a set of keywords that reflect the couple’s style) and the text descriptions that vendors put on their storefront. Recommendations are made by representing both of them as text embeddings and recommending venues that are close to the Style Quiz results in the embedding space.

The recommendations are further augmented by LLM-generated reasoning copy which highlights to our couples why a specific recommended venue matches their personal style. In this way, we can further personalize the results.

The screenshot below shows the venue recommendations in our product as a carousel that couples can browse, along with the reasoning for this example (the Brooklyn Botanical Garden).

We measured the impact of the semantic search-based recommendations quantitatively in an A/B test and found a 5.9% increase in conversion compared to the control group.