I have in the instructions, that my gpt-5 mini agent should include sources to the end of its response. It usually does, but sometimes not. The answers are literally incomplete, because sometimes it even acknowledges the links in the beginning of the response (like “Here are some links…”), but no links follow. I use bb-code for the links btw. Responses API.
Did anyone encounter this problem?
what API are you using, Completions or Responses?
Hi, thanks, I use Responses.
2 Likes
Are you getting response status “incomplete”?
2 Likes
No, I am not getting this status.
1 Like
Do you get better reliability from the full fat models (non-mini/nano)?
Honestly I have not yet tested this that much. Overally this is a tricky to debug, because it rarely says things like “Here are the links…” and no links then. It is usually just omitted sources, when can be attributed to non following the prompt well, but the rare occurences of explicitly mentioning the links hint that at least some and maybe all problems with missing links are a model problem, not a prompt one.
1 Like
I think sometimes one might want to design a supervisor layer that looks at the responses before sending them out and asks itself “does this response make sense” …