Use a LLM to train against many versions of metadata and subjects to suggest and draft a metadata for a portal or AGOL object based upon the key words used. Perhaps also use the summary statement in the initialization of the ChatGPT in creating the metadata summary. Allow users to edit further, allow LLM to further train on the enhancements made and corrections made. This would work perfectly on AGOL as well as portals with very narrow subject matters (medial geography, defense, etc.)
Also - Ask me about View-A-Tor....
