i have been using it for some time. Great, but it has a tendency to give short answers and don't use the full context available. even if you set the max response size to 20k.
I suppose they favored concise answers in the fine tunning process and reused the model to train the 32k.
Things that I have been doing: paste full table DDL with entities and other classes involved and asking for some refactoring of a class. Also that and some stacktraces. It definitely helps, but the price (at least on azure) make me think twice before using.
Also azure playground is awful.