The research assistant no longer does reasoning

fxgn

The research assistant no longer has a reasoning step, so the quality is much worse, and it often just starts reasoning in the response to the user

It should use a reasoning model

Rehan

Do you have any examples of the bad responses?

fxgn

EffortsFrom your thread doesn't seem to be available

Rehan sure, here's one example I had a while ago and just replicated now:

https://assistant.kagi.com/share/0f2f2d1b-2aa5-4246-b77a-3a1105708f88

The answer is fine, but as you can see, there are no reasoning steps, and instead the model just outputs regular text in between the intermediate steps, which should not be there. I don't have a comparison point but I strongly assume the lack of reasoning makes the answer worse, and this tracks with the vibes I've been getting from using research mode (I've almost stopped using it in the past weeks because of how useless it's become)

EffortsFrom

+1 regarding looking into degraded quality

@Rehan This I ran today, still on the old Assistant UI:
https://kagi.com/assistant/99e990c7-e3f9-43e5-bdfd-2c4295c22efd

Feel free to take a look at it. I vented my frustration in the thread (lol I know), but so you see what I pointed out.

Rehan

EffortsFrom Could you elaborate on what you did not like about the response? Could you give the non experimental research assistant a go and see if you like its responses better?

Ran your query through it: https://assistant.kagi.com/share/4bf0ce47-e0bb-44f4-9a05-bf54247b1f25

Rehan

fxgn The change there is more a visual change than anything. The intermediate text used to be hidden away in the reasoning blocks instead. This change is a consequence of other changes made to support a more agentic workflow in all other models.

It is hard for us to correct or improve off vibes. If there are instances where a response was not hitting the mark, please do report it!

fxgn

Rehan alright, here's an example of a specific issue - the quick mode seems to completely ignore my system prompt.

https://assistant.kagi.com/share/8f20708e-9db6-4020-90bc-0e1dfd4ac654

You can see my user instructions in the second message. They clearly say to avoid bold font, lists, and inline headings in paragraphs. And yet, the answer contains all of those things.

Seems to be due to the change of the Quick mode from Gemini to GLM, which is completely incapable of following formatting instructions and always produces slop AI-style answers.

(For comparison, a much better answer from Gemini: https://assistant.kagi.com/share/ea9c49fe-b149-4b53-93b1-c1e5d4f8d69b)