Model consults ResearchAssistant and never prints a response

artemvmin

Often times, the model prints out boilerplate text about it reaching out to ResearchAssistant, but never does and never prints an actual response. This happens with roughly 5% of my searches.

Give me an actual answer after consulting ResearchAssistant.
Don’t print the ResearchAssistant part at all. Sometimes it prints this boilerplate, then erases it, and replaces it with the fully resolved response. The boilerplate is distracting.

bitzbyte

Here’s another example of ResearchAssistant not returning any data.
https://kagi.com/assistant/622f8f92-b38d-469c-b9cb-a723a8516224

nichu42

I've read this might be Sonnet-related? Also using Sonnet as main LLM and this bug happens quite often.

paksmol

Assistant: Claude 3.5 Sonnet
Issue: Assistant repeats itself when failing to search online for more information.

Occured multiple times on different threads.

The assistant should've done a web search for the additional information it needs for the response and give the answer straightaway.

paksmol

happened again

nichu42

Same issue as https://kagifeedback.org/d/5117-model-consults-researchassistant-and-never-prints-a-response?

silvenga

Ishan512

Often times the assistant says it needs to search for information and calls research agent and then doesn't give any response

Thread: (version 4 is where it's stuck) https://kagi.com/assistant/1b515609-b544-493f-9496-bca88f4cc764?branch=01c529b7-ac4d-4a10-8cb0-4434307d1c79

Shouldn't get stuck. (Also I don't really need to see the system messages of calling agents etc, but that's a nit)

nichu42

Duplicate of https://kagifeedback.org/d/5117-model-consults-researchassistant-and-never-prints-a-response

Dannn404

Here is the chat where this happened. Note the two branches when I had it regenerate the answer, but got essentially the same answer.

I knew that iOS 17.4 was released, but wanted to know the date of release. But it just stated that it will search the web, but never did.

This is the first time I've encountered this.

I expected the assistant to search the web when asked, and ideally realize that it's information is out-of-date and get the release date as originally asked.

doggofan

When asking The Assistant Claude 3.5 Sonett with Web Access there is a chance it will respond with a message ending with what looks like it is talking to ResearchAssistant, then it exits without a useful response. Replying with kagi researchagent seems to put it back on track but it is a strange interaction.

My prompt was
kagi code assistant
because I was hoping it could tell me about the "code" option

The response was pretty unhelpful, but the ending was what was strange.

To provide you with accurate and up-to-date information about the Kagi Code Assistant, I'll need to consult the ResearchAgent. Let me do that for you.

I apologize, but I don't have specific information about a "Kagi code assistant." Kagi is known for its privacy-focused search engine, but I don't have confirmed details about a dedicated code assistant from them. To provide accurate information, I'll need to research this topic.

ResearchAgent, could you please find current information about any code assistant or coding-related tools offered by Kagi?

Desired functionality: when web access is available, the assistant should look up the query instead of posing it to the user. The assistant also should avoid messages that appear to address the user as ResearchAgent.

azdanov

Maybe another cool option would be to show some sort of a bubble or indicator that show a web search happened?

As an example, github copilot chat has this fold where it shows what it did. In this instance it looked at this file when asked to explain.

Maybe web access can be similar in some way?

azdanov

Like:

Summarized the query
Did the web search
Processed the results

silvenga

doggofan

silvenga similar issue. what surprised me most is it seems to think the user is "ResearchAgent" which I assume is part of the web access systems. This could confuse new users, and could be a concern if it indicates a problem with the initial prompt or filtering systems. Not sure prompts/agents are necessarily "secrets" but they shouldn't show up in normal usage.