According to the documentation, the retrieval tool when provided to the assistant should automatically provide relevant context from files to the model when its called. Custom GPTs are successfully doing this, Assistants are consistently not. To make matters worse, the Assistant is not using retrieval unless explicitly prompted. The GPT on the other hand just “knows” the information that was provided in its files.
I have been testing this pretty heavily as well and have found similar results. However, I think if you start to add more context in the instructions of what documents it has access to, what is in each of those documents, and their basic structure. I tested this by having multiple retrival documents and limiting the thread to one document out of all of them. I then posed a question that would be answered by one of the other documents without explicity saying so. It said the answer would more likely be in the other document and that the current document does not have the information needed.
I am still trying some different alternative Instructions, but the way you have to format for the Assistants API seems to be different than the Custom GPTs.