Kagi Assistant is extremely slow

er1z

Want to confirm. Getting a response from assistant takes much longer than previously. Different devices, fiber and LTE.

httpjames

It's extremely slow again. Reproducible in US and EU regions, so it's not isolated. Rendering of the web app and requests to the streamable endpoints are all slow/failing.

I was able to reproduce the backend timing out and not loading any threads at all:

httpjames

This issue needs to be prioritized. It should not be in the clarification status.

It's genuinely not a technical challenge to fetch 3k objects out of a database.

JayP

I just started using Kagi recently, and signed up For the Pro plan today. I started trying to use the assistant, with the plan of not renewing my perplexity subscription. But yes, it’s extremely slow in generating responses compared to anything else I’ve used. I want to love it and use it! But seems like a big downgrade unfortunately

Anonymous12

JayP are there any example prompts you can give to recreate it? I have not been noticing any issues with Assistant on my end today or lately. Possibly this could be something regional?

I would be happy to try testing on my end here on US-WEST.

er1z

Anonymous12 doesn't matter. If I switch model to Gemini, everything works noticeably faster.

httpjames

The latest update has removed slowness from prompting but thread_list is still unoptimized.

httpjames

thread_list has been paginated which fully solves slowness with listing threads. thread_open remains unoptimized and is unusably slow with high thread volume. Parsing the JSON data element in the /assistant HTML is a workaround.

httpjames

It's been 4 months and /thread_open remains unoptimized to the point of non functionality. No response despite providing tons of telemetry to prove it 😀

nicole5978

Hey @httpjames,

First off, thank you for the detailed reporting and patience. My apologies that this has dragged on as long as it has, the information you provided helps us narrow down the issue and is much appreciated.

Yes, we have implemented thread_list pagination and some prompting improvements to resolve the 2/3 issues described here, but thread_open poorly scaling is a gap we have yet to close. Significant improvements, including performance and scaling, are coming to Assistant in the near term, and I will post back here with an update as soon as I can. Thank you for your patience, and staying on this.

Nicole