AI 가드레일

from databricks.sdk import WorkspaceClient
from databricks.sdk.service.serving import (
    AiGatewayConfig,
    AiGatewayGuardrails,
    AiGatewayGuardrailParameters
)

w = WorkspaceClient()

w.serving_endpoints.put_ai_gateway(
    name="my-agent-endpoint",
    guardrails=AiGatewayGuardrails(
        input=AiGatewayGuardrailParameters(
            safety=True  # 안전 필터 활성화
        )
    )
)

w.serving_endpoints.put_ai_gateway(
    name="customer-support-agent",
    guardrails=AiGatewayGuardrails(
        input=AiGatewayGuardrailParameters(
            safety=True,
            valid_topics=[
                "customer support",
                "order inquiry",
                "product information",
                "shipping status",
                "return and refund"
            ],
            invalid_topics=[
                "politics",
                "religion",
                "competitor products",
                "investment advice",
                "medical advice"
            ]
        )
    )
)

w.serving_endpoints.put_ai_gateway(
    name="customer-support-agent",
    guardrails=AiGatewayGuardrails(
        output=AiGatewayGuardrailParameters(
            safety=True,
            pii={
                "behavior": "BLOCK"  # PII 감지 시 응답을 차단합니다
            }
        )
    )
)

w.serving_endpoints.put_ai_gateway(
    name="customer-support-agent",
    guardrails=AiGatewayGuardrails(
        output=AiGatewayGuardrailParameters(
            safety=True,
            pii={"behavior": "MASK"},
            invalid_keywords=[
                "경쟁사_이름",
                "내부기밀",
                "비공개"
            ]
        )
    )
)

from databricks.sdk.service.serving import AiGatewayRateLimit

w.serving_endpoints.put_ai_gateway(
    name="customer-support-agent",
    guardrails=AiGatewayGuardrails(
        input=AiGatewayGuardrailParameters(safety=True),
        output=AiGatewayGuardrailParameters(safety=True, pii={"behavior": "BLOCK"})
    ),
    rate_limits=[
        # 사용자별 분당 20회 제한
        AiGatewayRateLimit(
            calls=20,
            renewal_period="minute",
            key="user"
        ),
        # 전체 분당 1000회 제한
        AiGatewayRateLimit(
            calls=1000,
            renewal_period="minute",
            key="endpoint"
        )
    ]
)

from databricks.sdk import WorkspaceClient
from databricks.sdk.service.serving import (
    AiGatewayConfig,
    AiGatewayGuardrails,
    AiGatewayGuardrailParameters,
    AiGatewayRateLimit,
    AiGatewayUsageTrackingConfig
)

w = WorkspaceClient()

# 종합 가드레일 설정
w.serving_endpoints.put_ai_gateway(
    name="customer-support-agent",

    # 가드레일 설정
    guardrails=AiGatewayGuardrails(
        # 입력 필터
        input=AiGatewayGuardrailParameters(
            safety=True,
            valid_topics=["customer support", "order inquiry", "product info"],
            invalid_topics=["politics", "religion", "medical advice"]
        ),
        # 출력 필터
        output=AiGatewayGuardrailParameters(
            safety=True,
            pii={"behavior": "BLOCK"},
            invalid_keywords=["경쟁사A", "경쟁사B"]
        )
    ),

    # 속도 제한
    rate_limits=[
        AiGatewayRateLimit(calls=30, renewal_period="minute", key="user"),
        AiGatewayRateLimit(calls=2000, renewal_period="minute", key="endpoint")
    ],

    # 사용량 추적
    usage_tracking_config=AiGatewayUsageTrackingConfig(enabled=True)
)

-- 가드레일 차단 이벤트 분석
SELECT
    DATE(timestamp) AS date,
    COUNT(*) AS total_requests,
    COUNT(CASE WHEN guardrail_action = 'BLOCKED' THEN 1 END) AS blocked_count,
    ROUND(
        COUNT(CASE WHEN guardrail_action = 'BLOCKED' THEN 1 END) * 100.0 / COUNT(*), 2
    ) AS block_rate_pct
FROM catalog.schema.agent_logs_payload
GROUP BY DATE(timestamp)
ORDER BY date DESC;