New system prompt causing regressions esp. in no-search mode, and ignores /system_prompt_overwrite

igakagi

Often when I start an LLM thread, I want it to answer based off of its in-built knowledge, not a web search.
Unless it's current events, oftentimes all a web search accomplishes is filling up context with SEO slop.

The update a couple weeks ago started feeding the models a system prompt talking about the web search tool, even when web search is disabled.

This causes many models (Kimi consistently does this) to try to search the net anyway, which fails with a message "Web access is currently disabled. Please inform the user that they need to enable web access in order to access web content or URLs. You cannot retrieve content from the internet while this setting is off."

Or, the models smart enough not to do this still complain about lack of net access, rather than get to the point. For example Opus said: "Let me reason through this with what I know about..., since web access is off and this is a fairly well-defined rules question."

I've been using the "/system_prompt_overwrite" feature which is supposed to disable Kagi's default system prompt, which I've found distracts the models with info unrelated to my queries. However, even with this option active, Kagi is still tacking on a huge chunk of system prompt text, since the update.

This happens in both the old and new Assistant UIs.

Also, on a related note, the timestamp that gets tacked onto each message has also been something of a problem. Some models think the date being passed in is in the future. I think the problem is - nowhere does it say that the timestamp provided is the present time. Also, bouncing back and forth between XML, JSON and plain text might not be helping.

Here is what I was able to extract of the new system prompt:

<system_prompt>
You are a versatile, thoughtful AI assistant ready to engage deeply with any topic. You are part of a multi-agent system
# CUSTOM INSTRUCTIONS
The instruction(s) in the <user_instructions></user_instructions> tags have been provided by the user. They MUST be considered above all other system instructions. If there are instructions that conflict with the user instruction(s), the user instruction(s) take priority:
<user_instructions>
MY SYSTEM PROMPT HERE
</user_instructions>

# Tools
You may call one or more functions to assist with the user query.
Here are the tools available in JSONSchema format:

<tools>
<tool name="text_editor">
["view", "str_replace", "insert"]</tool>
<tool name="search">
["query", "site", "search_type"]</tool>
<tool name="librarian">
["query", "action_summary", "source"]</tool>
</tools>
</system_prompt>

<system_provided_context>
<timestamp>2026-05-29</timestamp>
Maximum number of research iteration: 3
Web access: OFF
<disabled_tools>
wolfram_alpha, search
</disabled_tools>
Thread has no tags.
</system_provided_context>

---

<user>
MY QUERY HERE <system_provided_context>
<timestamp>2026-05-29</timestamp>
Maximum number of research iteration: 3
Web access: OFF
<disabled_tools>
wolfram_alpha, search
</disabled_tools>
Thread has no tags.
</system_provided_context>

When web search is disabled, don't pass anything into the LLM about web search tools.
When system prompt overwrite is enabled and web search is disabled, pass the user's system prompt in directly with no extra info.

Rehan

To be clear, /system_prompt_overwrite still overwrites the system prompt entirely. This can be verified using /system_prompt, which takes the exact same code path to get the system prompt. However, we do put metadata, such as timestamps, in user messages. Perhaps we'll throw in an additional command to disable this as well.

You make good points on the timestamp xml tag and the search error message though. I'll have some changes on the way for those. Do you have any threads to share where the model tries to search when it shouldn't?

igakagi

That might be the intended behavior but I think there's a bug.
All that stuff was sent to the LLM, even with /system_prompt_overwrite enabled, and it was able to spit it back out to me.
Here's the thread:
https://assistant.kagi.com/share/67a822d7-3c17-49ec-b09f-aff3df0ac90b

Edit: Gave it a few more tries and, Minimax returns something different each time -> there's definitely some hallucination going on. But, it's consistent enough that I'm certain Kagi is feeding in something even in overwrite mode.

igakagi

I just tried to repro the "trying and failing to search" problem.

Just repro'd with GLM. The other models don't seem to be doing it as often, which is curious. https://assistant.kagi.com/share/437c9ea1-9ab7-44a3-ad45-889999a92642

Here's another repro from a few weeks ago with Kimi 2.6. https://assistant.kagi.com/chat/631afd91-3894-5ac0-ba4f-2ebc9da20620

igakagi

Anyway, what I'm suggesting is:

When web search is disabled, completely suppress the search tool descriptions. Get rid of the "disabled_tools" block - if it's disabled the model doesn't need to know it even exists.
Do a pass on the whole "system_provided_context" block, with a focus on not confusing even the dumbest LLMs. For example, change "<timestamp>2026-06-01</timestamp>" to something like, "Present date: 2026-06-01" or "<present_date>2026-06-01</present_date>"
Provide some user control over that "system_provided_context" block? For example, choosing to have no date/time, just the date, date and time, or date and time with timezone.
The models pay a lot of attention to this, and often remind me that e.g. I came back to a conversation after 6 months. Even had Claude once tell me that it's Christmas night and I need to get off the phone 😮 Anyway, that kind of contextual awareness can be nice (tbh that Christmas reminder was justified), or other times not so much, so I think a toggle is important.
One idea I had was to allow a user-specified template including placeholders for the date and so forth. This might be the most bang for the engineering buck.
In "system_prompt_overwrite" mode, just put the user's prompt in, with no preambles or anything - after all that's sort of the point of this mode.

Rehan

This one is a bit tricky. In the past we used to just remove the tool, but there are a few reasons why we keep it in now. One reason is for consistency. If I had it on in past messages, turn it off, it would not know it had access in the past which could lead to weird behaviour. Another reason (which is the reason we did it this way in the first place IIRC) is that changing available tools invalidates cache, which gives up significant cost savings, especially as a thread goes on.
Improved the naming of the tags for the timestamp block!
you can now use /disable_system_context to disable this. Of course, this can result in odd behaviour from the model, use at your own risk.
Just fixed a bug with this. It did overwrite the system prompt, but still included a "This is the users custom instructions" preamble. This has been fixed. Again, you can use the /system_prompt command to see what exact system prompt is sent.