track evaluator errors from sdk

[isahers1]

No description provided.

[hinthornw]

If we commit to doing this, we'll need to handle case where you return EvaluationResult as well

[isahers1]

thanks for calling this out. Tried to handle all the cases I could think of:

• return {"key":"foo","score":1} -> should extract ["foo"]
• return {"score":1} -> should extract name of function
• return EvaluationResult(...) should extract key from object
• return EvaluationResults(results=...) should extract keys from all the EvaluationResults objects passed to 'results'. This case is harder, but it works for simple invocations.