Why the model keeps referring to the "provided documents " in reply when using RAG

We’ve run into this issue again.

Previously, we were able to work around it by switching to gpt-5-mini, but now that model appears to be affected as well. We are currently moving to the newer gpt-5.4-mini, where the issue does not seem to occur.

That said, this is very frustrating. From our perspective, it feels like older models are silently regressing over time, which forces us to migrate to newer and more expensive models just to maintain the same behavior and output quality.

Could someone from OpenAI clarify whether there have been any changes to gpt-5-mini that could explain this regression, especially in RAG scenarios where the model starts explicitly referring to provided documents in the final answer?