The Superlinked team was out in full force at Berlin Buzzwords this year, stirring up plenty of conversation around vector search. They came prepared with hotel-search themed swag, handing out frisbees, sunglasses, and more sharkie stickers than anyone expected, all to promote the Superlinked booth. The giveaways brought people in, but the real draw was the conversations around where vector search is heading.

On the Monday, our Co-founder Ben and DevRel Filip delivered a talk on Superlinked’s Mixture of Encoders approach.

Instead of splitting structured data like filters and price ranges from unstructured data like text descriptions, Mixture of Encoders blends everything into one vector space. Each data type, whether text, images, categories or numerical values, has its own encoder. These specialized outputs are combined into a single embedding that captures the full meaning of a query.

This allows the system to handle complex searches like “family-friendly hotels near midtown under $200” in a single pass. The model understands both the semantics and the constraints directly, without requiring extra filters or reranking logic. The result is more accurate retrieval, simpler architecture and better performance.

‍

Superlinked has already seen this approach succeed in production, delivering millions of dollars in added revenue through better recommendations and discovery. They have also open-sourced datasets and demo implementations to help others try out the approach.

The team left Berlin with lots of positive feedback, plenty of new contacts and an even stronger buzz around Mixture of Encoders. If you want to watch the full talk, you can find it here:

‍

Did you grab a frisbee from us at Buzzwords? Let us know! If you missed us and still want to find out about our Mixture of Encoders approach in more depth, why not have a chat with us and see how we can help?