Skip to main content

Bedrock Guardrails

LiteLLM supports Bedrock guardrails via the Bedrock ApplyGuardrail API.

Quick Start​

1. Define Guardrails on your LiteLLM config.yaml​

Define your guardrails under the guardrails section

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: "bedrock-pre-guard"
    litellm_params:
      guardrail: bedrock  # supported values: "aporia", "bedrock", "lakera"
      mode: "during_call"
      guardrailIdentifier: ff6ujrregl1q      # your guardrail ID on bedrock
      guardrailVersion: "DRAFT"              # your guardrail version on bedrock
      aws_region_name: os.environ/AWS_REGION # region guardrail is defined
      aws_role_name: os.environ/AWS_ROLE_ARN # your role with permissions to use the guardrail

Supported values for mode​

  • pre_call Run before LLM call, on input
  • post_call Run after LLM call, on input & output
  • during_call Run during LLM call, on input Same as pre_call but runs in parallel as LLM call. Response not returned until guardrail check completes

2. Start LiteLLM Gateway​

litellm --config config.yaml --detailed_debug

3. Test request​

Langchain, OpenAI SDK Usage Examples

Expect this to fail since since ishaan@berri.ai in the request is PII

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": ["bedrock-pre-guard"]
  }'

Expected response on failure

{
  "error": {
    "message": {
      "error": "Violated guardrail policy",
      "bedrock_guardrail_response": {
        "action": "GUARDRAIL_INTERVENED",
        "assessments": [
          {
            "topicPolicy": {
              "topics": [
                {
                  "action": "BLOCKED",
                  "name": "Coffee",
                  "type": "DENY"
                }
              ]
            }
          }
        ],
        "blockedResponse": "Sorry, the model cannot answer this question. coffee guardrail applied ",
        "output": [
          {
            "text": "Sorry, the model cannot answer this question. coffee guardrail applied "
          }
        ],
        "outputs": [
          {
            "text": "Sorry, the model cannot answer this question. coffee guardrail applied "
          }
        ],
        "usage": {
          "contentPolicyUnits": 0,
          "contextualGroundingPolicyUnits": 0,
          "sensitiveInformationPolicyFreeUnits": 0,
          "sensitiveInformationPolicyUnits": 0,
          "topicPolicyUnits": 1,
          "wordPolicyUnits": 0
        }
      }
    },
    "type": "None",
    "param": "None",
    "code": "400"
  }
}

PII Masking with Bedrock Guardrails​

Bedrock guardrails support PII detection and masking capabilities. To enable this feature, you need to:

  1. Set mode to pre_call to run the guardrail check before the LLM call
  2. Enable masking by setting mask_request_content and/or mask_response_content to true

Here's how to configure it in your config.yaml:

litellm proxy config.yaml
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: "bedrock-pre-guard"
    litellm_params:
      guardrail: bedrock
      mode: "pre_call"  # Important: must use pre_call mode for masking
      guardrailIdentifier: wf0hkdb5x07f
      guardrailVersion: "DRAFT"
      aws_region_name: os.environ/AWS_REGION
      aws_role_name: os.environ/AWS_ROLE_ARN
      mask_request_content: true    # Enable masking in user requests
      mask_response_content: true   # Enable masking in model responses

With this configuration, when the bedrock guardrail intervenes, litellm will read the masked output from the guardrail and send it to the model.

Example Usage​

When enabled, PII will be automatically masked in the text. For example, if a user sends:

My email is john.doe@example.com and my phone number is 555-123-4567

The text sent to the model might be masked as:

My email is [EMAIL] and my phone number is [PHONE_NUMBER]

This helps protect sensitive information while still allowing the model to understand the context of the request.