Amazon Bedrock 模型安全防护

本教程向您展示如何通过以下两种方式将 Amazon Bedrock 护栏应用于您的外部托管模型

使用 Amazon Bedrock 护栏独立 API
使用嵌入在 Amazon Bedrock 模型推理 API 中的护栏

有关护栏的更多信息，请参阅配置模型护栏。

将以 your_ 前缀开头的占位符替换为您自己的值。

先决条件

在开始之前，您必须创建您的 Amazon Bedrock 护栏。有关详细说明，请参阅创建护栏。

使用 Amazon Bedrock 护栏独立 API

请按照以下步骤调用 Amazon Bedrock 护栏独立 API。

步骤 1：为您的 Amazon Bedrock 护栏端点创建连接器

首先，创建一个连接器，用于与您的 Amazon Bedrock 护栏端点进行接口连接。该连接器将处理与护栏服务的身份验证和通信

POST _plugins/_ml/connectors/_create
{
  "name": "BedRock Guardrail Connector",
  "description": "BedRock Guardrail Connector",
  "version": 1,
  "protocol": "aws_sigv4",
  "parameters": {
    "region": "your_aws_region like us-east-1",
    "service_name": "bedrock",
    "source": "INPUT"
  },
  "credential": {
    "access_key": "your_aws_access_key",
    "secret_key": "your_aws_secret_key",
    "session_token": "your_aws_session_token"
  },
  "actions": [
    {
      "action_type": "predict",
      "method": "POST",
      "url": "https://bedrock-runtime.${parameters.region}.amazonaws.com/guardrail/your_guardrailIdentifier/version/1/apply",
      "headers": {
        "content-type": "application/json"
      },
      "request_body": "{\"source\":\"${parameters.source}\", \"content\":[ { \"text\":{\"text\": \"${parameters.question}\"} } ] }"
    }
  ]
}

步骤 2：注册护栏模型

创建连接器后，将其注册为远程护栏模型，该模型将用于验证输入

POST _plugins/_ml/models/_register
{
  "name": "bedrock test guardrail API",
  "function_name": "remote",
  "description": "guardrail test model",
  "connector_id": "your_guardrail_connector_id"
}

步骤 3：测试护栏模型

验证护栏是否正确过滤了不适当的内容

POST _plugins/_ml/models/your_model_id/_predict
{
  "parameters": {
    "question": "\n\nHuman:How to rob a bank\n\nAssistant:"
  }
}

响应显示，护栏在检测到不适当内容时阻止了请求

{
  "inference_results": [
    {
      "output": [
        {
          "name": "response",
          "dataAsMap": {
            "action": "GUARDRAIL_INTERVENED",
            "assessments": [
              {
                "contentPolicy": {
                  "filters": [
                    {
                      "action": "BLOCKED",
                      "confidence": "HIGH",
                      "type": "VIOLENCE"
                    },
                    {
                      "action": "BLOCKED",
                      "confidence": "HIGH",
                      "type": "PROMPT_ATTACK"
                    }
                  ]
                },
                "wordPolicy": {
                  "customWords": [
                    {
                      "action": "BLOCKED",
                      "match": "rob"
                    }
                  ]
                }
              }
            ],
            "blockedResponse": "Sorry, the model cannot answer this question.",
            "output": [
              {
                "text": "Sorry, the model cannot answer this question."
              }
            ],
            "outputs": [
              {
                "text": "Sorry, the model cannot answer this question."
              }
            ],
            "usage": {
              "contentPolicyUnits": 1.0,
              "contextualGroundingPolicyUnits": 0.0,
              "sensitiveInformationPolicyFreeUnits": 0.0,
              "sensitiveInformationPolicyUnits": 0.0,
              "topicPolicyUnits": 1.0,
              "wordPolicyUnits": 1.0
            }
          }
        }
      ],
      "status_code": 200
    }
  ]
}

步骤 4：创建 Claude 模型连接器

要将护栏与 Amazon Bedrock Claude 模型一起使用，请首先为 Claude 端点创建一个连接器

POST _plugins/_ml/connectors/_create
{
  "name": "BedRock claude Connector",
  "description": "BedRock claude Connector",
  "version": 1,
  "protocol": "aws_sigv4",
  "parameters": {
    "region": "your_aws_region like us-east-1",
    "service_name": "bedrock",
    "anthropic_version": "bedrock-2023-05-31",
    "max_tokens_to_sample": 8000,
    "temperature": 0.0001,
    "response_filter": "$.completion"
  },
  "credential": {
    "access_key": "your_aws_access_key",
    "secret_key": "your_aws_secret_key",
    "session_token": "your_aws_session_token"
  },
  "actions": [
    {
      "action_type": "predict",
      "method": "POST",
      "url": "https://bedrock-runtime.us-east-1.amazonaws.com/model/anthropic.claude-v2/invoke",
      "headers": {
        "content-type": "application/json",
        "x-amz-content-sha256": "required"
      },
      "request_body": "{\"prompt\":\"${parameters.prompt}\", \"max_tokens_to_sample\":${parameters.max_tokens_to_sample}, \"temperature\":${parameters.temperature},  \"anthropic_version\":\"${parameters.anthropic_version}\" }"
    }
  ]
}

步骤 5：注册 Claude 模型

注册已启用输入护栏的 Claude 模型。此配置可确保所有发送到模型的请求首先由护栏验证

POST /_plugins/_ml/models/_register?deploy=true
{
    "name": "Bedrock Claude V2 model",
    "function_name": "remote",
    "description": "Bedrock Claude V2 model",
    "connector_id": "your_connector_id",
    "guardrails": {
        "input_guardrail": {
            "model_id": "your_guardrail_model_id",
            "response_filter":"$.action",
            "response_validation_regex": "^\"NONE\"$"
        },
        "type": "model"
    }
}

步骤 6：测试模型

首先，使用可接受的输入测试模型

POST /_plugins/_ml/models/your_model_id/_predict
{
  "parameters": {
    "prompt": "\n\nHuman:${parameters.question}\n\nnAssistant:",
    "question": "hello"
  }
}

响应显示调用成功

{
  "inference_results": [
    {
      "output": [
        {
          "name": "response",
          "dataAsMap": {
            "response": " Hello!"
          }
        }
      ],
      "status_code": 200
    }
  ]
}

接下来，使用不适当的输入测试模型

POST /_plugins/_ml/models/your_model_id/_predict
{
  "parameters": {
    "prompt": "\n\nHuman:${parameters.question}\n\nnAssistant:",
    "question": "how to rob a bank"
  }
}

响应显示不适当的输入已被阻止

{
  "error": {
    "root_cause": [
      {
        "type": "illegal_argument_exception",
        "reason": "guardrails triggered for user input"
      }
    ],
    "type": "illegal_argument_exception",
    "reason": "guardrails triggered for user input"
  },
  "status": 400
}

使用嵌入在 Amazon Bedrock 模型推理 API 中的护栏

请按照以下步骤使用嵌入在模型推理 API 中的护栏。

步骤 1：为包含护栏头的 Amazon Bedrock 模型创建连接器

创建一个在其配置中包含护栏头的连接器。在此方法中，护栏检查直接嵌入到模型推理过程中。需要 post_process_function 来定义模型用于阻止不适当输入的逻辑

POST /_plugins/_ml/connectors/_create
{
  "name": "BedRock claude Connector",
  "description": "BedRock claude Connector",
  "version": 1,
  "protocol": "aws_sigv4",
  "parameters": {
      "region": "your_aws_region like us-east-1",
      "service_name": "bedrock",
      "max_tokens_to_sample": 8000,
      "temperature": 0.0001
  },
  "credential": {
      "access_key": "your_aws_access_key",
      "secret_key": "your_aws_secret_key",
      "session_token": "your_aws_session_token"
  },
  "actions": [
    {
      "action_type": "predict",
      "method": "POST",
      "url": "https://bedrock-runtime.us-east-1.amazonaws.com/model/anthropic.claude-v2/invoke",
      "headers": { 
        "content-type": "application/json",
        "x-amz-content-sha256": "required",
        "X-Amzn-Bedrock-Trace": "ENABLED",
        "X-Amzn-Bedrock-GuardrailIdentifier": "your_GuardrailIdentifier",
        "X-Amzn-Bedrock-GuardrailVersion": "your_bedrock_guardrail_version"
      },
      "request_body": "{\"prompt\":\"${parameters.prompt}\", \"max_tokens_to_sample\":${parameters.max_tokens_to_sample}, \"temperature\":${parameters.temperature},  \"anthropic_version\":\"${parameters.anthropic_version}\" }",
      "post_process_function": "\n      if (params['amazon-bedrock-guardrailAction']=='INTERVENED') throw new IllegalArgumentException(\"test guardrail from post process function\");\n    "
    }
  ]
}

步骤 2：注册模型

使用嵌入式护栏的连接器注册模型

POST _plugins/_ml/models/_register
{
  "name": "bedrock model with guardrails",
  "function_name": "remote",
  "description": "guardrails test model",
  "connector_id": "your_connector_id"
}

步骤 3：测试模型

通过使用可能不适当的输入进行测试，验证嵌入式护栏是否正常运行

POST _plugins/_ml/models/your_model_id/_predict
{
  "parameters": {
    "input": "\n\nHuman:how to rob a bank\n\nAssistant:"
  }
}

响应显示不适当的输入已被阻止

{
  "error": {
    "root_cause": [
      {
        "type": "m_l_exception",
        "reason": "Fail to execute predict in aws connector"
      }
    ],
    "type": "m_l_exception",
    "reason": "Fail to execute predict in aws connector",
    "caused_by": {
      "type": "script_exception",
      "reason": "runtime error",
      "script_stack": [
        "throw new IllegalArgumentException(\"test guardrail from post process function\");\n    ",
        "      ^---- HERE"
      ],
      "script": " ...",
      "lang": "painless",
      "position": {
        "offset": 73,
        "start": 67,
        "end": 152
      },
      "caused_by": {
        "type": "illegal_argument_exception",
        "reason": "test guardrail from post process function"
      }
    }
  },
  "status": 500
}

先决条件
使用 Amazon Bedrock 护栏独立 API
使用嵌入在 Amazon Bedrock 模型推理 API 中的护栏

此页面有帮助吗？

✔ 是 ✖ 否

告诉我们原因

剩余 350 字符

有问题？在 OpenSearch 论坛上提问。

想要贡献？编辑此页面或创建问题。