- Issue created by @andrewbelcher
- 🇩🇪Germany marcus_johansson
So, #1 is tested and works, but its not great to change if someone is actaully using the raw output and the objects for their business logic. However the module is not stable yet, so we could do that. OpenAI, FireworksAI and Ollama does exactly that at the moment.
I think #4 is a good way to go, but for all operation types for 2.0 release. We currently log the rawoutput for any operation type, but should instead always have an array version of the output.
We could also do a mix where we add the properites now, and its optional for providers to implement and then the logging method checks if the property is set, otherwise it tries to use the raw output. And then make it required in the interface for 2.0 release. We are thinking about doing the same thing for output tokens handling before 2.0.