How to define an LLM-as-a-judge evaluator

Key concepts

• LLM-as-a-judge evaluator

LLM applications can be difficult to evaluate because often they're generating conversational text for which there's no single "right" answer. An imperfect but valuable way to evaluate such applications is to use a second LLM to judge the outputs of the first. This can be especially useful if a smaller model is used in the application and a larger, better model is used for evaluation.

Custom evaluator via SDK

For maximal control of evaluator logic, we can write a custom evaluator and run it using the SDK.

Python

Requires langsmith>=0.1.145

from langsmith import evaluate, traceable, wrappers, Client
from openai import OpenAI
# Assumes you've installed pydantic
from pydantic import BaseModel

# Optionally wrap the OpenAI client to trace all model calls.
oai_client = wrappers.wrap_openai(OpenAI())
  
def valid_reasoning(inputs: dict, outputs: dict) -> bool:
  """Use an LLM to judge if the reasoning and the answer are consistent."""

  instructions = """\

Given the following question, answer, and reasoning, determine if the reasoning \
for the answer is logically valid and consistent with question and the answer.\
"""

  class Response(BaseModel):
    reasoning_is_valid: bool

  msg = f"Question: {inputs['question']}\nAnswer: {outputs['answer']}\nReasoning: {outputs['reasoning']}"
  response = oai_client.beta.chat.completions.parse(
    model="gpt-4o",
    messages=[{"role": "system", "content": instructions,}, {"role": "user", "content": msg}],
    response_format=Response
  )
  return response.choices[0].message.parsed.reasoning_is_valid

# Optionally add the 'traceable' decorator to trace the inputs/outputs of this function.
@traceable
def dummy_app(inputs: dict) -> dict:
  return {"answer": "hmm i'm not sure", "reasoning": "i didn't understand the question"}

ls_client = Client()
questions = ["how will the universe end", "are we alone"]
dataset = ls_client.create_dataset("big questions")
ls_client.create_examples(dataset_id=dataset.id, inputs=[{"question": q} for q in questions])

results = evaluate(
  dummy_app,
  data=dataset,
  evaluators=[valid_reasoning]
)

See here for more on how to write a custom evaluator.

Prebuilt evaluator via langchain

See here for how to use prebuilt evaluators from langchain.