How do teams plan out search encoder improvements in absence of click data?

Is the new GPL(Generative Pseudo Labeling) training working out for you?

We used GPL in the past, but it’s not as common nowadays (unless Cohere / OpenAI are implementing it, but they have the resources to manually annotate so it seems unlikely). If you are still interested in GPL check out this article.

Now that we have LLMs we can actually get models like GPT-4 and co to do the dataset annotation for us