microsoft
Privacy masking Agent (like a key manager) extension (?)
I'd like to be able to do tasks related to private data such as a bank account using a smarter LLM(on remote). Agents may write code to request web data, however I wouldn't want to send the output of the web reply to the bank directly to a remote LLM. Instead, I'd like to have a (not necessarily very smart) local LLM-mechanism that filters out such data as I'll explain below.
A local LLM that looks at text data and derives simple regex'es of possibly sensivite values and then gives it to a regex engine for substitution on the text.
I think I can do it myself by writing an extension but I need some suggestions. The "privacy masking agent" needs to work like a key manager that protects the data. The main reason I ask is, a chatroom architecture may not be suitable for such an agent, so I need some guidance whether it could be possible using TaskWeaver, and request some pointers to how?
