Need some pointers on training for specific domain and different customers

I’m exploring the possible use of a conversational interface on the domain of InfoSec risk identification.

Thoughts:

  1. You can prompt ChatGPT to take on a certain role, like: “you are an infosec consultant specialised in working with non-IT people”; How would this work with Rasa? Where would I need to configure that?
  2. Different customers (i.e. organisations) tend to differ in for instance their systems landscape, so in conversations they will use the same product names within an organization but different product names between organisations. Could I have like different ‘pools’ or something where the domain knowledge is the same but the model is sensitive to specific naming within a pool?
  3. Security issue: how can I prevent information about risks and vulnerabilities from one client organization doesn’t leak to others?

Also pointers to general introductory content to how this training of these models would work are greatly appreciated.