Embedding Models

An embedding model is what gives your search a sense of meaning. It reads each record and turns it into a list of numbers that captures what the record is about, so that things which mean similar things end up near each other. With one attached, a search for “running shoes” can also surface “trainers”, and “laptop for travel” can find a “lightweight ultrabook” — matches with no words in common. Without one, search matches words; with one, it matches meaning. You bring the model, with your own API key.

Auto-plays · use Back / Next to step through at your own pace.

Add an embedding model

1. Go to Models → Embedding Models and click Add Embedding Model.
2. Pick a provider — OpenAI, Gemini, Voyage or Microsoft Foundry. Embedding runs on your own key: your models, your limits, your bill.
3. Choose the model and paste your provider API key. The key is verified with a live call before it is saved.
4. Click Add. The model is now available to attach to lists and groups.

Attach it — and what “semantic” unlocks

Choose an embedding model when you create a list (or add one to a group later). Once records are embedded, three things become possible that plain keyword search can’t do:

• Meaning-based matching — find records by what they mean, not just the words they contain.
• Related results — “more like this” on a detail page, powered by which records sit closest in meaning.
• Image and multimodal search — with a model that embeds images, search a list by a picture as well as by text.

Changing a group’s model happens in the background

Every record has to be re-embedded when the model changes, which takes time on a large list. So on a group the change is handled as a background job: you pick the new model and search keeps running on the old vectors until the new ones are ready, then switches over. There is no downtime and no half-migrated state — the switch is all-or-nothing, once the new embeddings have caught up.

A worked example

Say you run a recipe site. With plain search, “quick weeknight dinner” only finds recipes that literally say “quick” or “weeknight”. Attach an embedding model and the same search surfaces a 20-minute stir-fry and a one-pan pasta that never use those words but clearly fit — because the model understands that’s what the phrase means. Your records didn’t change; search just got a sense of what they’re about.