How do you proper personalize search & recommendations?

Hi everyone! I built a Python SDK for search and recommendations, but not in the way you would expect. I’d be interested in how people here would evaluate the approach.

It uses BehaviorGPT, a 12.5B parameters behavioral model trained on sequences of user actions to predict what happens next. Note; this isn’t a language model.

Products are embedded from their images and attributes, and the model uses purchase patterns learned during pre-training to rank items in a new catalog. You can use it for natural-language search and recommendations conditioned on intent and interaction history.

The SDK, getting-started notebook, and research links are here:
github.com/Unbox-AI/behaviorgpt

It uses our free hosted API and requires an API key. The current catalog limit is 20,000 items.

Would be interesting to know where you would expect this approach would thrive, but also to struggle, for example, niche catalogs or behavior that differs from the pretraining data?

If you try the notebook, feedback on setup or unclear instructions would also be welcome.