Auto select suitable LLM based on user interaction when they start Kagi assistant

marcel001122

mixchiev I’m not suggesting suggestions to be shown. But categories.

Code
Analyze
Summarize
Create

Something like that. So users aren’t using Deepseek R1 to just summarise their 5 page long document when 4o mini probably can do the task quicker and cheaper.

dps

marcel001122

dps Yes ditto. But have a UI element to this scenario when pro-actively user has been asked to choose what's the purpose and then AI makes the choice.

dps

marcel001122 I would personally prefer Kagi determining this for me, but obviously that's going to be more work for them. This feature would be a good first step toward a more comprehensive recommendation like above!

marcel001122

dps Kagi could do it for us. They have benchmark data and they test LLM all the time so would know which ones perform the best for what 'type' of query. Then this can be fine tuned over time.

RoxyRoxyRoxy

marcel001122 Only issue I can see with that approach is an increase in overhead per prompt. Seems to me like they’d need to run a separate LLM to classify the category of a prompt.

dps

marcel001122 By Kagi determining this for me, I meant them recommending which "category" a thread would fall in. I think your idea would be easier to implement though and could be a solid feature addition

marcel001122

RoxyRoxyRoxy or pass queries to pre-decided LLM with ..if..then.. logic?

RoxyRoxyRoxy

marcel001122 Oh my bad lmfao, I totally misunderstood what you meant. 😅 Yes you’re right; when you said “Kagi does it for us” I thought you meant selecting the model without any user intervention, just putting in the prompt and hitting go.

dps

Perplexity and Grok have both auto as the default model type. Taking care of deciding what functions/features a user needs by their query.

marcel001122

dps Yes something similar but instead of dropdown, if there was a tap/touch/click that UI element is just visible. Here I would barely be inclined to change. Plus these are one word descriptors. I like what is in Grok UI but not as model selector, instead 'draw a lone astronaut..' type of cards.

It would have: Today I want to....
[Draw] [C0de] [Learn] [Reason] [Summarise] [Proofread] etc. User would click on whatever and that LLM will be ready in background. Now that ready LLM doesn't have to be a single one. Kagi could choose to have 2-3 alternatives ready and select based on actual query that comes in. But at least Deepseek R1 won't be used to learn about how to jailbreak an iPhone for example.

Luis

We are currently experimenting with this. We'll have some updates to share soon.

dreifach

I'd like this to be an on/off toggle function. I don't always want the model chosen for me.