Send user IP address to Assistant LLM/WA

Zambyte

The AI model queries Wolfram Alpha for the answer to the question "Where am I?" which presumably is answered with respect to the language model, rather than the end user.

The answer should accurately respond with the location of the user.

Vlad

Zambyte That would involve sending user's IP address to the API/LLM, something we are not doing now.

KagiForMe

Vlad sending user's IP address to the API/LLM

That is a big step. I feel like one of the selling Kagi features is the usage of AI in a somewhat private form (no IP, no learning, ..). I am not sure whether implementing this would be worth it. If it is, I feel like a toggle would be needed for your privacy-focused user base to feel comfortable using the Kagi AI features at all.

Zambyte

@Vlad The reason why I submitted this as a bug rather than as a feature is because it confidently says the wrong information. Another valid way to resolve this issue would be to simply tell the user that the request is outside the current capability of the system. Ideally I think it should give the right information, but at the very least it should not give the wrong information.

Vlad

Zambyte LLM does not know that it is giving you right or wrong information. Asking it where you are is probably not a question an LLM should be asked.

Zambyte

Vlad This example is slightly contrived, but it's the most basic example of location based question that someone might ask.

Another example is comparing the results of searching "weather" using the web search with the reply to "What is the weather today?".

The former gives the correct information in the users current location. The latter lets the user know what the weather is in Washington D.C.

More generally: Queries related to location should not use the wrong location. They should either use the right location, or refuse to reply with a statement about the limitations of the system.

Zambyte

Also to be clear: I'm not asking for a general "if the response isn't true, don't reply with it". Obviously a language model can't do that. I'm suggesting a prompt be added that tells the model to reject queries that require the location of the user (ie: "What is the weather?" would be rejected, but "what is the weather in NYC?" would be properly processed, because it doesn't require knowledge of the users location). This would be a minimal fix to the bug that I reported.

A potential feature that would also fix the bug would be to get the users city when asking a location based question, and answering based on that (ie: a user who asks "What is the weather?" while in NYC will have their question transformed into "What is the weather in NYC?" before it is passed to Wolfram).