Hey guys,
GPT-4o search preview in our app doesn’t show or cite sources anymore. The model is clearly more up-to-date than the standard GPT-4o, for example when you ask for the news of the day. But the answers seem random, still sometimes out of date, and it stopped citing sources even when explicitly asked to do so. No website names, no links to websites. It worked last week. Any idea why this is happening?
Can confirm there’s an issue with citations on OpenAI’s end. We’re working on a fix now. Sorry for the issues.
Thank you! Do you have any tips in general on how to improve the sourcing? Compared to ChatGPT it appears to be worse. Is it because of the licensing?
Hm! It’s the same model behind search in ChatGPT. Have you tried fiddling with your system prompt to experiment with different outputs?
The web search tool has internal instructions on producing the in-document citations, that on Responses, are received positionally.
The web search output is highly-patterned, a recitation based directly on instructions out of your control.
However, you can get a bit of control back by demanding that use of web search results also must have footnote citations added after the response, with directly-clickable markdown links. The ideal place for this instruction would be the tool itself, but OpenAI has not allowed customization of internal tools that would satisfy a great number of applications correctly. Therefore, you must make it by system message.
Counterpoint:
chatgpt-4o-latest that is supposed to be similar giving no tool use.Not API search results:
So far I haven’t seen any results from established news sources/publishers. I suspected that the reason for this is that OpenAI licenses this content for ChatGPT, not for OpenAI’s APIs. Maybe I’m wrong.
