Azure AI Evaluation SDK for Python

使用内置评估器和自定义评估器，评估生成式 AI 应用的性能。

安装

pip install azure-ai-evaluation

# 启用远程评估支持
pip install azure-ai-evaluation[remote]

环境变量

# 用于 AI 辅助型评估器
AZURE_OPENAI_ENDPOINT=https://.openai.azure.com
AZURE_OPENAI_API_KEY=
AZURE_OPENAI_DEPLOYMENT=gpt-4o-mini

# 用于 Foundry 项目集成
AIPROJECT_CONNECTION_STRING=

内置评估器

质量评估器（AI 辅助型）

from azure.ai.evaluation import (
    GroundednessEvaluator,
    RelevanceEvaluator,
    CoherenceEvaluator,
    FluencyEvaluator,
    SimilarityEvaluator,
    RetrievalEvaluator
)

# 使用 Azure OpenAI 模型配置初始化
model_config = {
    "azure_endpoint": os.environ["AZURE_OPENAI_ENDPOINT"],
    "api_key": os.environ["AZURE_OPENAI_API_KEY"],
    "azure_deployment": os.environ["AZURE_OPENAI_DEPLOYMENT"]
}

groundedness = GroundednessEvaluator(model_config)
relevance = RelevanceEvaluator(model_config)
coherence = CoherenceEvaluator(model_config)

质量评估器（基于 NLP）

from azure.ai.evaluation import (
    F1ScoreEvaluator,
    RougeScoreEvaluator,
    BleuScoreEvaluator,
    GleuScoreEvaluator,
    MeteorScoreEvaluator
)

f1 = F1ScoreEvaluator()
rouge = RougeScoreEvaluator()
bleu = BleuScoreEvaluator()

安全评估器

from azure.ai.evaluation import (
    ViolenceEvaluator,
    SexualEvaluator,
    SelfHarmEvaluator,
    HateUnfairnessEvaluator,
    IndirectAttackEvaluator,
    ProtectedMaterialEvaluator
)

violence = ViolenceEvaluator(azure_ai_project=project_scope)
sexual = SexualEvaluator(azure_ai_project=project_scope)

单行评估

from azure.ai.evaluation import GroundednessEvaluator

groundedness = GroundednessEvaluator(model_config)

result = groundedness(
    query="What is Azure AI?",
    context="Azure AI is Microsoft's AI platform...",
    response="Azure AI provides AI services and tools."
)

print(f"Groundedness score: {result['groundedness']}")
print(f"Reason: {result['groundedness_reason']}")

使用 evaluate() 进行批量评估

from azure.ai.evaluation import evaluate

result = evaluate(
    data="test_data.jsonl",
    evaluators={
        "groundedness": groundedness,
        "relevance": relevance,
        "coherence": coherence
    },
    evaluator_config={
        "default": {
            "column_mapping": {
                "query": "${data.query}",
                "context": "${data.context}",
                "response": "${data.response}"
            }
        }
    }
)

print(result["metrics"])

复合评估器

from azure.ai.evaluation import QAEvaluator, ContentSafetyEvaluator

# 一站式涵盖全部质量指标
qa_evaluator = QAEvaluator(model_config)

# 一站式涵盖全部安全指标
safety_evaluator = ContentSafetyEvaluator(azure_ai_project=project_scope)

result = evaluate(
    data="data.jsonl",
    evaluators={
        "qa": qa_evaluator,
        "content_safety": safety_evaluator
    }
)

评估应用目标（Application Target）

from azure.ai.evaluation import evaluate
from my_app import chat_app  # 你的应用

result = evaluate(
    data="queries.jsonl",
    target=chat_app,  # 接收 query 并返回 response 的可调用对象
    evaluators={
        "groundedness": groundedness
    },
    evaluator_config={
        "default": {
            "column_mapping": {
                "query": "${data.query}",
                "context": "${outputs.context}",
                "response": "${outputs.response}"
            }
        }
    }
)

自定义评估器

基于代码的评估器

from azure.ai.evaluation import evaluator

@evaluator
def word_count_evaluator(response: str) -> dict:
    return {"word_count": len(response.split())}

# 在 evaluate() 中使用
result = evaluate(
    data="data.jsonl",
    evaluators={"word_count": word_count_evaluator}
)

基于提示词的评估器

from azure.ai.evaluation import PromptChatTarget

class CustomEvaluator:
    def __init__(self, model_config):
        self.model = PromptChatTarget(model_config)
    
    def __call__(self, query: str, response: str) -> dict:
        prompt = f"Rate this response 1-5: Query: {query}, Response: {response}"
        result = self.model.send_prompt(prompt)
        return {"custom_score": int(result)}

记录至 Foundry 项目

from azure.ai.projects import AIProjectClient
from azure.identity import DefaultAzureCredential

project = AIProjectClient.from_connection_string(
    conn_str=os.environ["AIPROJECT_CONNECTION_STRING"],
    credential=DefaultAzureCredential()
)

result = evaluate(
    data="data.jsonl",
    evaluators={"groundedness": groundedness},
    azure_ai_project=project.scope  # 将结果记录至 Foundry
)

print(f"View results: {result['studio_url']}")

评估器参考表

最佳实践

1. 优先使用复合评估器，实现全面评估
2. 正确映射列名 —— 列名不匹配将导致静默失败
3. 记录至 Foundry，便于跨运行跟踪与对比
4. 创建自定义评估器，满足领域特定指标需求
5. 当拥有标准答案（ground truth）时，优先使用 NLP 评估器
6. 安全评估器需传入 Azure AI 项目作用域（azure_ai_project scope）
7. 批量评估比逐行循环更高效

参考文件