Hey, im making a project which includes using LLM to act as "search engine"

I need LLM to use tool calling to request for which category of products to search from with pipeline:
Category (LLM gets all main categories)
LLM picks sub category
LLM picks another sub category

This may not be the best way but im expecting like 500 categories in total so i cannot just give them all to LLM since it will eat too many tokens.

Maybe there is better way which im open to aswell, i was thinking of embedding but i tried and it seems to be really bad with connecting "slang" words for categories/products.