Options in the Assistant to change the “Thinking” Effort and Search “Depth” of models.

fs1010

Many reasoning AI models have different levels of reasoning “effort” they can expend before answering, such as Grok 3 mini (Low/High Thinking). Having an option to change that would be useful. Alongside this, being able to tell the models how “deep” you want to research would be useful. Sometimes you explicitly ask a model to do a search/use ResearchAgent and it just… doesn’t. Having a more reliable way to affect that would be nice.

For the Thinking Effort option, I would expect something like this brain icon with the respective options per model
For search depth, maybe there could be a “Deep Research” mode button ala ChatGPT, or (more experimental) maybe a slider of sorts that lets you pick roughly how many sources you want referenced (like 1-10, 10-50, 50-100).

clementpoiret

On Kagi Assistant, you either choose between thinking or non thinking models. What would be great instead would be to have a new toggle for compatible model, responsible for setting thinking level.

As an example, I select o3, a new toggle appears to right with thinking modes "low, mid, high". This would declutter the the model selection list (e.g., no more separated Claude 4 Sonnet and Claude 4 Sonnet Thinking), and increase the user's control over a model's behavior

Example:
Example

fs1010

I created a very similar issue and it’s marked as under review.
https://kagifeedback.org/d/7353-options-in-the-assistant-to-change-the-thinking-effort-and-search-depth-of-models

clementpoiret

Ohh, awesome. Didn't saw yours, sorry 🙂

yokoffing

Theo just released commentary on this very subject:

watch 14:26-15:36

TL;DR: "Low" or "Medium" are appropriate. "High" just makes the request 3x more expensive.

In other words: The quality of the output changes very little, but the cost increase is substantial.

@Vlad If that's true, then Assistant could have a toggle for reasoning (like we do with Assistant + web search). The toggle to activate reasoning would be the equivalent of "Low" or "Medium" mode in other LLM providers so that we're not burning tokens. Assistant could also have a separate setting for reasoning by default under https://kagi.com/settings/assistant > Custom Assistants.

Would love to see this ASAP. We don't even have a toggle to activate "Low" reasoning for a model as cheap as Gemini 2.5 Flash, and I'd never use Claude 4 Opus as that's expensive and overkill for my use case.

Article in the video discussing the future of subscription-based LLM services: https://ethanding.substack.com/p/ai-subscriptions-get-short-squeezed

rudyfink

To the extent Kagi is applying or using API settings for the different LLMs in the assistant, it would be helpful to see those configurations. I do not mean API keys or anything private, just the configuration flags / settings. Using LLMs can, sometimes, be an exercise in trying to understand if there is some existing setting / context that is not visible that is creating an issue. With visibility into what (if any) settings Kagi is applying, some of those issues could be more easily debugged by the user. If Kagi is not applying any settings, no worries, that makes the blurb for that LLMs API all the easier 🙂.

I was imagining something like a settings "gear" icon or a small question mark "?" icon added to the model descriptions in the Assistant drop-down. Right now that has a short description of the model and information on price and accuracy. I think that is great! The thought would just be some way to surface additional information if there is more to know.

fs1010

rudyfink

To follow up on this point. Kagi engineering was good enough to look into the issue that raised this suggestion (thanks again). The issue was, apparently, an AI hallucination claiming “document processing not enabled.” I add that as another point that this feature could have use as users navigate our fun future of the tool itself pretending it can or cannot do something 🙂.