[BUG] Web Search Button stays grey on OpenAI 4.1 models and o4 mini

[DoS007]

Environment

GitHub v2-dev branch

Description

Using the newest version of the v2-dev branch (updated today)

https://platform.openai.com/docs/models/gpt-4.1
https://platform.openai.com/docs/models/o4-mini

These require responses api support, but that was added as i can now see and use https://platform.openai.com/docs/models/o4-mini-deep-research which is only available in the responses api.

Edit: Also on o3 which has web search support: https://platform.openai.com/docs/models/o3

Device and browser

Brave 1.80.115 Chromium: 138.0.7204.97 (Offizieller Build) (64-Bit) on Ubuntu 24 LTS

Screenshots and more

No response

Willingness to Contribute

• 🙋‍♂️ Yes, I would like to contribute a fix.

[DoS007]

Just like to add: Via playground on openai the websearch works, but it's different to chatgpt search at least for o3 and o4-mini.

O3 web search with big context (specifically for the web search) on scientific publications were 5 publications and 3 on medium. And one (!) search request costed ~70 cent (euro). All publications are read with the named top models. With chatgpt search there probably is a less costly model used for reading (maybe 4.1 nano, 4o-mini or 4.1 mini) but thus can read more and is less costly if you do more than 1 o3 search request (e.g. for scientific papers) per day.

So I'm not as hyped for the web search functionality as I was.

[enricoros]

Hi @DoS007 thanks for the feedback. The "web" button you see there is to load web pages (to markdown) when the "browser" tool is configured in settings. If you use the Big-AGI I provide in the website I pre-configure it.

• The "web" will fetch a web page, transform to markdown, screenshot (if wanted) and attach it to the current message.
• OpenAI has a few models that have search embedded (and are supported) as well as some where search needs to be used as a "tool"
• Also we will deploy MCP tools including Search, so there's another option

What's your ideal use for the web search: everyday context retrieval from the web, deep-search-like, or native search integration for the models that have it supported?

One more thing: Grok 4 supports search natively for web and X posts, as another avenue to source grounding information.

[DoS007]

@enricoros

Ah, ok (am using an own instance via vercel).

I assume option one is only for a single page (but maybe possible for more using google api for search).

(Assessing the three points) I think for most people 2. and 3. will be most interesting, because copy and paste is probably for many people as good as option 1. And second will be also important as the search functionality may perform better if the search functionality usage is learned with reinforcement learning (maybe see e.g. Grok). This maybe could mean that when using search functionality, it's best to use the out-of-the-box integrated one on which the models are optimized .

(Using official search) The thing is, search on openai with o4-mini is free. So that use case is not as important. For me ideal use would be a really grounded answer, but I think that will have some complexity (especially using the last argument of the last paragraph). So native search integreation for the models that have supported it would be still the most relevant (that the search functionality of openai on o4-mini and o3 isn't good at the moment is not a problem of big-agi and may change in the future).

(Custom search) Of cource it would be nice to have some custom search functionality, but there might be some difficulty to get it right (model-with-prompt for planing search terms e.g. o3, model-with-prompt for evaluating relevance e.g. o3, model-with-prompt writing the most important things to the question 4.1-mini or 4.1-nano, model-with-prompt give an answer with the excerpts of the searches e.g. o3).

(Important apart that) Apart from the above points, I also think that a saveable custom system prompt (and setting as default) may currently have higher priority for most people I would assume (but maybe it's only me). E.g. o3 with custom prompt maybe could lead to nicer output formatting and so on.