Superlinked had a full showing at HaystackConf 2025, one of the top conferences for developers building search, retrieval, and discovery systems. We weren’t just attending. We sponsored a dinner, gave a lightning talk, and even saw our name pop up on a keynote slide titled “Glimmers of Hope.”

The Talk: Metadata-Aware Vector Embeddings and Mixture of Encoders

Daniel delivered a lightning talk introducing a new approach to embedding design. The concept? Mixture of Encoders, like Mixture of Experts, but purpose-built for the complex structured and unstructured data that powers real-world search.

Daniel explained that most enterprise data doesn’t come in neat text chunks. It’s a mix of product descriptions, prices, categories, ratings, images, user histories, and geographic info. Instead of forcing all of that into a single encoder, Superlinked uses multiple specialized encoders. One might handle text, another works on numbers, another on categories or locations. These are combined into a single embedding space that reflects the full shape of the data.

He also highlighted a major issue with standard text embeddings. They fail at handling numbers. A model that treats "3.7" the same way it treats "banana" isn't going to give you good search. Superlinked solves this by giving numerical and structured values their own embedding logic, preserving their meaning and boosting retrieval quality​.

The talk was short but hit the mark. Daniel said the idea landed immediately. The audience followed along, asked sharp questions, and clearly related to the challenge of dealing with messy, multi-field data.

One of the best moments? Spotting Superlinked listed on a keynote slide by John Berryman under a roundup of emerging solutions trying to bridge structured and unstructured search

The Dinner: Community Vibes and Real Talk

Daniel and Ben also hosted a Superlinked dinner for conference attendees. It turned into one of the most memorable parts of the event. No sales pitches, no slides. Just honest conversations with engineers, researchers, and builders who are deep in the weeds of modern retrieval.

People shared what’s working, what’s breaking, and what they wish existed. The energy was real. The food was good. The community feel? Absolutely priceless.

Massive thanks to the HaystackConf organizers for pulling off such a high-quality event. And huge thanks to everyone who came to our dinner. If you're exploring smarter embeddings or wrestling with metadata-rich search, we’d love to hear from you.

Check out Vectorhub if you'd like hands-on tutorials for your vector-based projects.

‍