This behaviour appears to be new (may as of today?) - I am seeing completely different output depending on whether the assistant’s output is being streamed or not. This can be reproduced in the “Assistant Playground”.
When streaming, the results are accurate and detailed. When non-streamed, the results are generally innacurate, often hallucinated, and very often fail to correctly refer to the documents. Either the documents are ignored or they are misquoted.
Is anyone else experiencing this?