Kagi Assistant, aside from running LLMs locally, is about as private as interactions with LLMs can get for most user/consumers (when thinking about exposure to API/LLM providers alone). Kagi's contracts with API providers act as a shield and its retention policies are straightforward.
That said, one of the key privacy weaknesses in terms of API providers being able to identify and theoretically profile the queries submitted by Kagi's individual users in aggregate is custom instructions (and notes, being a beta feature).
Having a universal custom instruction (in my case, for example, specifying a British English output in a particular way) or custom assistant-specific custom instructions is about as identifying as it gets. The promises contained in contracts with API providers aside, it remains a possibility that users could theoretically be profiled. Big tech is about as unscrupulous as it gets, so this isn't far off reality given recent LLM breaches/controversies.
I thereby propose a type of Privacy Pass analogue for the Assistant: for a new thread, the ability to toggle a "Privacy Pass" (or whatever it could/would be called) that disables/strips all custom instructions and notes from the query and the thread history.
This is substantively different from #1931 (Add incognito mode for Search Assistant to allow use on confidential content) given this is more about privacy from API providers rather than privacy from Kagi itself.
