Deep Eval和Confident AI

价格

链接

Free套餐：基本够用

Starter套餐：20美元/月

Quick Start

官方

最新版本

pip install -U deepeval

注册 Confident AI

获取api key

登陆

deepeval login --api-key xxxx

创建文件test_example.py

from deepeval import assert_test
from deepeval.test_case import LLMTestCase
from deepeval.metrics import AnswerRelevancyMetric

def test_answer_relevancy():
    answer_relevancy_metric = AnswerRelevancyMetric(threshold=0.5)
    test_case = LLMTestCase(
        input="What if these shoes don't fit?",
        # Replace this with the actual output of your LLM application
        actual_output="We offer a 30-day full refund at no extra cost."
    )
    assert_test(test_case, [answer_relevancy_metric])

文件名必须是test_开头

运行

deepeval test run test_example.py

设置保存结果的文件夹

设置环境变量

export DEEPEVAL_RESULTS_FOLDER="./deep-eval-results"

deepeval test run的参数

并行

deepeval test run test_example.py -n 4

缓存

deepeval test run test_example.py -c

重复

deepeval test run test_example.py -r 2

钩子

...

@deepeval.on_test_run_end
def function_to_be_called_after_test_run():
    print("Test finished!")

基本概念

官方文档

Test Case	包含input/actual_output/retrieval_context
Dataset	Test Case的集合
Golden	相比 test case，少了 actual_output

一个实际的Test Case

test_case = LLMTestCase(
  input="Who is the current president of the United States of America?",
  actual_output="Joe Biden",
  retrieval_context=["Joe Biden serves as the current president of America."]
)

使用 Pytest 进行评估

from deepeval import assert_test
from deepeval.test_case import LLMTestCase
from deepeval.metrics import AnswerRelevancyMetric

dataset = EvaluationDataset(test_cases=[...])

@pytest.mark.parametrize(
    "test_case",
    dataset,
)
def test_customer_chatbot(test_case: LLMTestCase):
    answer_relevancy_metric = AnswerRelevancyMetric()
    assert_test(test_case, [answer_relevancy_metric])

@pytest.mark.parametrize 是 Pytest 提供的装饰器。它只是循环逐 EvaluationDataset 一评估每个测试用例。

不用CLI运行

文档

# A hypothetical LLM application example
import chatbot
from deepeval import evaluate
from deepeval.metrics import HallucinationMetric
from deepeval.test_case import LLMTestCase

# ……
test_cases = [first_test_case, second_test_case]

metric = HallucinationMetric(threshold=0.7)
evaluate(test_cases, [metric])

Test Case

一个标准的Test Case

test_case = LLMTestCase(
    input="What if these shoes don't fit?", #必选
    expected_output="You're eligible for a 30 day refund at no extra cost.", #必选
    actual_output="We offer a 30-day full refund at no extra cost.",
  
    context=["All customers are eligible for a 30 day full refund at no extra cost."], # 参考值
    retrieval_context=["Only shoes can be refunded."], # 实际检索结果
  
    latency=10.0
)

• context 是给定输入的理想检索结果，通常来自评估数据集，
• retrieval_context LLM应用程序的实际检索结果。

Dataset

手动生成dataset，并推送到Confident AI

from deepeval.test_case import LLMTestCase
from deepeval.dataset import EvaluationDataset

# 原始数据
original_dataset = [
    {
        "input": "What are your operating hours?",
        "actual_output": "...",
        "context": [
            "Our company operates from 10 AM to 6 PM, Monday to Friday.",
            "We are closed on weekends and public holidays.",
            "Our customer service is available 24/7.",
        ],
    },
    {
        "input": "Do you offer free shipping?",
        "actual_output": "...",
        "expected_output": "Yes, we offer free shipping on orders over $50.",
    },
    {
        "input": "What is your return policy?",
        "actual_output": "...",
    },
]

# 遍历，将生成 LLMTestCase 实例
test_cases = []
for datapoint in original_dataset:
    input = datapoint.get("input", None)
    actual_output = datapoint.get("actual_output", None)
    expected_output = datapoint.get("expected_output", None)
    context = datapoint.get("context", None)

    test_case = LLMTestCase(
        input=input,
        actual_output=actual_output,
        expected_output=expected_output,
        context=context,
    )
    test_cases.append(test_case)

# 将 LLMTestCase 数组变成 EvaluationDataset
dataset = EvaluationDataset(test_cases=test_cases)

# 推送到Confident AI
dataset.push(alias="My Confident Dataset")

链接

查看结果

在Confident AI的Web UI手动创建

支持中文

手动修改/Users/xxx/anaconda3/envs/LI311-b/lib/python3.11/site-packages/deepeval/synthesizer/template.py下的所有Prompt，增加

6. `Rewritten Input` should be in Chinse.

Metric

评估的维度

评估指标	描述
正确性和语义相似度	生成的答案 与 参考答案 的对比
Context Relevancy	查询 与 检索到的上下文 的相关性
Faithfulness	生成的答案 与 检索到的上下文 的一致性
Answer Relevancy	生成的答案 与 查询 的相关性

Confident AI