When I prompt the “gpt-4o-mini-search-preview” to search for “the latest sport news stories from last week (week 21) in Sweden” it returns 4 true and 6 false events, but then completely makes up the rest of the response, ending it saying “some of the above results are made up”.
I have tried to add instructions like “Do NOT make up any stories” etc., but nothing seems to work. Anyone else that have stumbled upon this problem?
You may have better search results if you use the web_search_options → user_location location API parameter for the AI model.
You can use system prompting to tell the model what it MUST do and get better following than telling what it MUST NOT do. There is no sense of “don’t make up stuff”, because to the AI model, all language is predicted and “made up”.
For web search, add today’s date to the system message. Add: “for answering events after your knowledge cutoff, you must directly repeat information only from web search results, or report that there is no information retrieved”.
The web search model has its own techniques and internal instructions and injection outside of any tool calling, so you can’t do much to guide the output.