Prompt flow evaluators

Introduction

Evaluators are custom or prebuilt promptflow flows that are designed to measure the quality of the outputs from language models.

Usage

Users can create evaluator runs on the local machine as shown in the example below:

import os
from pprint import pprint

from promptflow.core import AzureOpenAIModelConfiguration
from promptflow.evals.evaluate import evaluate
from promptflow.evals.evaluators import RelevanceEvaluator
from promptflow.evals.evaluators.content_safety import ViolenceEvaluator


def answer_length(answer, **kwargs):
    return {"value": len(answer)}


if __name__ == "__main__":
    # Built-in evaluators
    # Initialize Azure OpenAI Connection
    model_config = AzureOpenAIModelConfiguration(
        azure_endpoint=os.environ.get("AZURE_OPENAI_ENDPOINT"),
        api_key=os.environ.get("AZURE_OPENAI_KEY"),
        azure_deployment=os.environ.get("AZURE_OPENAI_DEPLOYMENT"),
    )

    # Initialzing Relevance Evaluator
    relevance_eval = RelevanceEvaluator(model_config)

    # Running Relevance Evaluator on single input row
    relevance_score = relevance_eval(
        answer="The Alpine Explorer Tent is the most waterproof.",
        context="From the our product list,"
        " the alpine explorer tent is the most waterproof."
        " The Adventure Dining Table has higher weight.",
        question="Which tent is the most waterproof?",
    )

    pprint(relevance_score)
    # {'gpt_relevance': 5.0}

    # Content Safety Evaluator

    # Initialize Project Scope
    project_scope = {
        "subscription_id": "e0fd569c-e34a-4249-8c24-e8d723c7f054",
        "resource_group_name": "rg-test",
        "project_name": "project-test",
    }

    violence_eval = ViolenceEvaluator(project_scope)
    violence_score = violence_eval(question="What is the capital of France?", answer="Paris.")
    pprint(violence_score)
    # {'violence': 'Very low',
    # 'violence_reason': "The system's response is a straightforward factual answer "
    #                    'to a geography question. There is no violent content or '
    #                    'language present.',
    # 'violence_score': 0}

    # Code based evaluator
    answer_length("The Alpine Explorer Tent is the most waterproof.")
    # {'value': 48}

    # Using multiple evaluators together using `Evaluate` API

    result = evaluate(
        data="evaluate_test_data.jsonl",
        evaluators={
            "answer_length": answer_length,
            "violence": violence_eval,
        },
    )

    pprint(result)