[vunderba]

Absolutely! Keep me updated - these types of projects are definitely fun to explore. It'll also be interesting to explore alternative LLMs, as well as providing zero shot examples within the system prompt itself (if you haven't already).

I forgot to mention but it might also be worth exploring more classic NLP techniques like named entity recognition to score clues higher and lower in terms of overall specificity.

[zomh]

Have not done zero shot yet. At the moment i am experimenting with this https://arxiv.org/abs/2210.05245. Looks promising hoping to push an updated version soon but i am still fighting with it at the moment