Unnecessary bold formatting in Assistant responses

voidpointer

When I use Kagi Assistant for creative work, using the Claude 3.7 model, it often inserts bold formatting in the prompt when I did not ask for it. This is unusual, because when I use my direct API usage of Claude 3.7 using Librechat, it does not do this. I can only assume there's some system prompt or other behavior customized for Assistant that is causing undesired artifacts in prompt responses. Screenshot below showing an example in the response output where I instructed AI to generate a sci-fi story.

My expectation is that the random formatting inconsistencies be removed.

yiwei-1

Have you tried adding custom instructions to ask it stop using bold formatting in its responses?

voidpointer

yiwei-1 I haven't, but part of what I'm saying is that I don't feel like I should have to. That's not the normal behavior for Claude 3.7...

RoxyRoxyRoxy

voidpointer Hi, can you confirm whether or not this is still occurring, given the upgrade to Claude 4?

voidpointer

Has there been some change by Kagi that I need to verify?

To directly answer your question, no, I haven't tested. As a workaround, I modified the system prompt to exclude bolding. I'd have to remove that and perform multiple tests to try to reproduce the issue. So far, I haven't been updated on the status of any root cause analysis performed or gotten confirmation from the Kagi team that they are actually causing this issue intentionally. There's just not enough information all around.

I am happy to test again but I want to make sure it's in response to an actual change or fix by the Kagi team, since I have limited time to test this.

yiwei-1

voidpointer

Sorry for the delayed update. We plan to expose options for users to overwrite most (if not all) of the current system prompt that we added for formatting, tool calling, etc. so that the behavior can be more fully customized.