[RFC] Add LLMEvaluator to create LLM-as-a-judge evaluators

[agola11]

It's currently quite annoying to use LLM-as-a-judge evaluators in code, and there is a bit of a disconnect between the SDK and the UI.

Our off-the-shelf evaluators don't even use tool calling

In the UI, you can specify the prompt and output schema. With LangSmith, you have to use .with_structured_output within a custom function. This can be a lot of boilerplate for the user.

Additionally, a JSON schema or generic pydantic model is likely not the best interface for allowing people to specify the score format for their LLM evaluators. Opted for something more opinionated, ContinuousScoreConfig and CategoricalScoreConfig

Important detail: I map each score to a tool as opposed to each argument of a tool. This allows other attributes, like explanation to be extracted and mapped to the same feedback entry.

Future work:

• Allow people to load these from a file
• Create off-the-shelf evaluators based off of LLMEvaluator
• async

[hinthornw]

Really like the idea of having an opinionated, easy evaluator def using tool calling