Setup Web Search

[mayoralink]

Hey there!

I'm absolutely new to this and I was following a guide on reddit but the post got deleted idk why. But anyways.

I have downloaded my models, right not primarily using Gemma4, but I am trying to enable it to be able to do web searches when connected to the internet. The settings where it defines the protocols and websites looks fine, but it is not able to actually perform web searches.

Any ideas on how to enable this? Thanks in advance.

[pabl1too]

I have the same problem, thought docker was the issue but honestly im lost right now since its my first LLM experience. Are you using docker as well?

[mayoralink]

Update: It's odd. Qwen3.5 is actually able to do WebSearches but Gemma 4 is not? Like huh?

[pabl1too]

i tested 3 different ollama models and none of them worked, tried to troubleshoot with claude but no luck either.. I wonder if a clean manual installation would fix it

[tkwok34]

I am new. But for some reason Gemma can't do web search (with duckduckgo) but Gwen 2 can. Not too sure where to start comprehending this.

[mayoralink]

I am new. But for some reason Gemma can't do web search (with duckduckgo) but Gwen 2 can. Not too sure where to start comprehending this.

Funnily enough, it is also Gemma 4 for me that is not being able to make searches, but Qwen 3.5 can? Like?

[TenshiKage]

I don't know if this will be useful but I ran in to an issue last night where web search stopped working. I am also using Gemma4.

When I originally set Odysseus up, I installed Natively. Then I tried switching to Docker without clearing out the native install first. I thought nothing of it as I could switch between Docker and Native and my main issue was getting tool calling to work with Gemma4 at the time.

Last night I tried doing a deep research task and it fell over because it could not pull sources.

This morning when I tried looking in to it, I ran the update-windows.bat file in the odysseus directory and it complained that it could not find the docker image for SearXNG.

I then loaded up docker (just Docker Desktop - I didnt start the container) and I then re-ran the update-windows.bat file and this time it was able to update SearXNG.

After this (still with Docker Desktop open) I was able to complete the deep research task I set off successfully.

As a test, I then closed down Docker Desktop and in Agent Mode I just asked it to do a web search on some stuff and it still worked. So I wonder if I messed something up when I tried to switch it running to Docker? I have never used Docker before so I am not sure but I hope this can help you.

[Skipxl]

Every video of people I see setting up Odysseus online the search function works out of the box, but for me it's not working either.

[itsmekaeiron]

Try changing from SearXNG to DuckDuckGo in Search section at the Settings page.

[mayoralink]

The problem I don't think is the method or search engine, Gemma 4 is the one that does not want to search, but Qwen3.5 can

[tkwok34]

sometime I find that using Duckduckgo, it does the search, but the model isn't taking the information and give false info back, and switching to Tavily, it works fine.

[Ashish-24-80]

I just found out that gemma 4 e4b can search the web .
You can try pressing the websearch button again.

Initially, it was unable to call the websearch even though web search was on.
I think the problem was in the indication that websearch was on.
Because sometimes the websearch might seem to be on, when in reality it is off.
It doesnt occur that much but i have seen it happen some times.

[Michael679089]

I downloaded gemme 4 e4b like you said but it didn't allow me to search the web.

My prompt was "search in the internet how to live life"

and it gave me back an empty response.

[1edwin1]

I believe the issue is dependent on the model when using external tools. for example I ran the issue with qwen3:30b using the search tool, but when using gpt-oss:20b it worked. I then made a custom system prompt for qwen in the prompt option and put this persona as qwen agent:

(copy starting here)

TOOL CALL FORMAT — mandatory. To invoke any tool, output a markdown fenced code block whose language tag is the tool name, containing JSON arguments. Example:

{"query": "example query", "max_results": 5}

Rules:

• Never use <tool_call> or </tool_call> tags.
• Never write tool calls as plain text, YAML, or function-call syntax like web_search(...).
• Never describe or announce a search — emit the fenced block and stop; the result will be provided to you.
• After receiving tool results, answer the user based on them. If another tool call is needed, use the same fenced-block format again.

(copy ends here)

try this and it should work :)

[Michael679089]

Where is this system prompt, i can't find it in the settings.

[SirThinksAl0t]

I believe the issue is dependent on the model when using external tools.

Yes, happened to me too.
For me, many models failed to do that, I was only successful with: gemma-4-e4b or gemma-4-26b-a4b-qat.

I think this is one of the reasons why PewDiePie is making a "Odysseus model".