Implement "personally identifiable information" remover

dreifach

It would be great to see a PII-remover, à la OpenAI Privacy Filter, built into Kagi Assistant by default. An option to enable/disable this, embedded within settings for greater customisation, would be a plus too.

An implementation of this kind would greatly improve the privacy performance of the Assistant, and address user concerns as raised elsewhere: #9921 (Custom instruction/note Assistant incognito mode)

An additional feature that scrambles custom instructions, whether within custom assistants or Assistant-wide custom instructions, would add to this too.

doskagi

I am in agreement with this. It would be very valuable for my use cases.

empower5069

Agreed, though I would prefer if the model was hosted on Kagi's servers rather than using an OpenAI API implementation to preserve privacy of the inputted data. It is a small model (only 1.5bn parameters) so I imagine this should be fairly straightforward from an infrastructure POV?

empower5069

Noting the (hopefully) incoming desktop app at https://kagifeedback.org/d/4766-assistant-as-a-dedicated-desktop-app

Ideally the model could be locally hosted within the desktop app to keep the input text on-device, before the output is sent to whichever model via API is being used.