After having used roboflow’s workflow creator for a few months now. I make all of my workflows with custom blocks in them, and have needed help with the defintions, attributes, or outputs of roboflow’s functions. Having gone to the ai assistant for help several times now, I have run into several problems I would like to see fixed:
-
Almost always, after 1-5 message exchanges with the assistant, the send message button will suddenly become unresponsive and I can’t send messages with the enter button either. It’s seemingly random as to when this happens, but when it does, I have to reload the page to be able to send messages again, which is annoying because it wipes the context of our conversation for reference.
-
When I ask it a question about a bug I’m having in a custom block (or any block for that matter), and it immediately goes to implement the fix it suggests in its response, but the auto-fix is screwed up. Usually the fix will delete all of my custom blocks, add an existing block in their place, and the block it adds will not do at all what I wanted my workflow to do. After it does this, I have to click “undo changes” and implement what it suggested myself (if the suggestions even work, see 3).
-
The assistant doesn’t seem to have complete knowledge regarding definitions, attributes, or outputs of roboflow’s functions. For example, when I asked it to give me more details about the output data from an object detections model’s predictions, it gives me the following response:
The code the assistant gives doesn’t work. It was only after searching through the inference developer guide and finding the definition for the object_detection_prediction datatype did I get my question answered truthfully and specifically (bounding box coordinates are accessed through a .xyxy list of tuples and everything else from a separate .data dictionary attribute). The assistant will have this sort of vague knowledge about anything I ask it regarding custom blocks. It will also change its mind about what’s correct or what causes error in code depending on what the error message is.
I know that the ai’s still in beta, but I would like to this implemented as a solid feature in the future, particulary for custom block development, if these problems are fixed.