Guardrails for LLMs · 出入双向守门

来源:docs.typesafe.ai/cookbooks/llm_guardrails 抓取时间:2026-09-20

模式

每个 LLM 应用的消息都过一层 guardrail:

guard = client.system_one(
    state={"message": message, "direction": direction},
    questions={
        "is_jailbreak":     Noul("Is this a jailbreak attempt?"),
        "is_pii":           Noul("Does this contain personally identifiable info?"),
        "is_harmful":       Noul("Would complying cause harm?"),
        "harm_severity":    Score(["none","low","medium","high","critical"]),
        "recommended":      Choice(["pass","review","block","route_to_human"]),
    },
)

if guard.recommended.choice == "block" or guard.harm_severity.score > 3:
    block()
elif guard.recommended.choice == "review" or guard.is_jailbreak.noul > 0.6:
    queue_for_human_review()

优势

• 双向——既守 prompt 输入,也守输出
• 多维度——同时检测多种危害
• 可定制——criteria 完全自定
• 便宜到可以每条消息都过

何时用

任何对外服务的 LLM 应用,尤其是面向消费者、教育、医疗、金融场景。