AI Models
Every service that searches or recommends products relies on one or more AI embedding models — labeled Nets in Panel. A Net converts text, images, or both into a vector; two products (or a query and a product) are considered similar when their vectors are close together. This page explains what each Net actually does, so the weights you set on a campaign’s Semantic Search section mean something concrete.
Turning Nets on
Section titled “Turning Nets on”Nets are activated per site under Site → AI methods in Panel — see Panel Configuration. A campaign can only be weighted toward a Net that’s already turned on for the site.
The Nets
Section titled “The Nets”-
🍵 Darjeeling Second Flush Net
Text-to-text. Understanding query meaning and matching it against product text (title, description). Also powers query understanding — extracting brand, size, color, and price signals from what a shopper typed.
-
🌋 Volcanic Water Net
Text-to-image, image-to-image. A general-purpose, high-fidelity model that works across both text and visual matching — a strong default for most Text Search and Visual Search setups.
-
🍇 Berry Punch Net
Text-to-image, image-to-image. Similar range to Volcanic Water, tuned to be lighter and faster — a good choice when result speed matters more than maximum precision.
-
🎨 Kiba Net
Image-to-image. A small, fast model for pure visual similarity — grouped under Color in Panel and commonly used for color-aware matching.
-
☕ Barraquito Net
Image-to-text. Also grouped under Color in Panel, alongside Kiba — enabled per site via the same AI methods panel.
How weights work
Section titled “How weights work”Each service campaign has separate weight sliders for text embeddings (product weights) and image embeddings (image weights). Turning on more than one Net and giving each a weight blends them: the final closeness between a query and a product is a weighted average across every active Net. Setting a Net’s weight to zero (or leaving it off) excludes it entirely.
There’s no single right combination — a text-heavy catalog with detailed descriptions benefits from a higher Darjeeling weight, while a highly visual catalog (fashion, home goods) benefits from leaning on Volcanic Water or Berry Punch. If you’re not sure where to start, Panel’s default configuration option when creating a campaign picks a sensible starting combination for you.
Where this shows up
Section titled “Where this shows up”- Text Search, Visual Search, Similar Products, and Cross-Selling campaigns all have a Semantic Search section for weighting Nets.
- Product Data uses AI models to predict attributes like color, but that’s a separate, unweighted prediction step rather than a campaign-configurable Net mix.