Skip to main content

Lakera AI

快速开始​

1. 在您的LiteLLM config.yaml中定义Guardrails​

在guardrails部分定义您的guardrails

model_list:
  - model_name: gpt-3.5-turbo
    litellm_params:
      model: openai/gpt-3.5-turbo
      api_key: os.environ/OPENAI_API_KEY

guardrails:
  - guardrail_name: "lakera-guard"
    litellm_params:
      guardrail: lakera  # 支持的值: "aporia", "bedrock", "lakera"
      mode: "during_call"
      api_key: os.environ/LAKERA_API_KEY
      api_base: os.environ/LAKERA_API_BASE
  - guardrail_name: "lakera-pre-guard"
    litellm_params:
      guardrail: lakera  # 支持的值: "aporia", "bedrock", "lakera"
      mode: "pre_call"
      api_key: os.environ/LAKERA_API_KEY
      api_base: os.environ/LAKERA_API_BASE

mode支持的值​

  • pre_call 在LLM调用之前运行，针对输入
  • post_call 在LLM调用之后运行，针对输入和输出
  • during_call 在LLM调用期间运行，针对输入 与pre_call相同，但在LLM调用期间并行运行。直到guardrail检查完成才返回响应

2. 启动LiteLLM网关​

litellm --config config.yaml --detailed_debug

3. 测试请求​

Langchain, OpenAI SDK 使用示例

预期这次调用会失败，因为请求中的ishaan@berri.ai是PII信息

curl -i http://localhost:4000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-npnwjPQciVRok5yNZgKmFQ" \
  -d '{
    "model": "gpt-3.5-turbo",
    "messages": [
      {"role": "user", "content": "hi my email is ishaan@berri.ai"}
    ],
    "guardrails": ["lakera-guard"]
  }'

预期失败的响应

{
 "error": {
   "message": {
     "error": "违反内容安全策略",
     "lakera_ai_response": {
       "model": "lakera-guard-1",
       "results": [
         {
           "categories": {
             "prompt_injection": true,
             "jailbreak": false
           },
           "category_scores": {
             "prompt_injection": 0.999,
             "jailbreak": 0.0
           },
           "flagged": true,
           "payload": {}
         }
       ],
       "dev_info": {
         "git_revision": "cb163444",
         "git_timestamp": "2024-08-19T16:00:28+02:00",
         "version": "1.3.53"
       }
     }
   },
   "type": "None",
   "param": "None",
   "code": "400"
 }
}

高级设置​

设置基于类别的阈值​

Lakera有两个用于prompt_injection攻击的类别：

  • jailbreak
  • prompt_injection
model_list:
  - model_name: fake-openai-endpoint
    litellm_params:
      model: openai/fake
      api_key: fake-key
      api_base: https://exampleopenaiendpoint-production.up.railway.app/

guardrails:
  - guardrail_name: "lakera-guard"
    litellm_params:
      guardrail: lakera  # 支持的值: "aporia", "bedrock", "lakera"
      mode: "during_call"
      api_key: os.environ/LAKERA_API_KEY
      api_base: os.environ/LAKERA_API_BASE
      category_thresholds:
        prompt_injection: 0.1
        jailbreak: 0.1
优云智算