Guardrails

Guardrails are AI safety controls that prevent the Context API from generating harmful, offensive, or inappropriate content. This page explains which actions are covered by guardrails, what types of content trigger violations, and what a guardrail error response looks like.

Understanding Guardrails in Knowledge Enrichment

To help ensure safe and responsible use of AI, the Hyland AI Platform has introduced guardrails in the Knowledge Enrichment service. These guardrails are designed to protect users from generating or receiving harmful, offensive, or inappropriate content.

Guardrails are built-in safety mechanisms that monitor and restrict certain types of content. They help prevent the generation of responses that may include:

• Sexual content
• Hate speech
• Insults
• Misconduct

These protections apply automatically to most available actions, such as summarizing text, classifying content, or generating descriptions.

note

Generating embeddings is not covered by guardrails.

Importance of Guardrails

Guardrails help maintain a respectful and safe environment for all users. They ensure that the AI behaves responsibly and aligns with community standards. If a request violates these guardrails, the system blocks the response and notifies the user.

Overview of Guardrail Configuration

The following table outlines the recommended guardrail levels for each task within Knowledge Enrichment:

Task Type	Applied Guardrails
Generate Description	Sexual Content, Hate Speech, Insults, Misconduct
Summarization	Sexual Content, Hate Speech, Insults, Misconduct
Classification	Sexual Content, Hate Speech, Insults, Misconduct
Named Entity Extraction	Sexual Content, Hate Speech, Insults, Misconduct
Metadata Extraction	Sexual Content, Hate Speech, Insults, Misconduct
Classify Image into Categories	Sexual Content, Hate Speech, Insults, Misconduct
Generate Image Description	Sexual Content, Hate Speech, Insults, Misconduct
Generate and Match Image Metadata	Sexual Content, Hate Speech, Insults, Misconduct
Detect Entities in Images	Sexual Content, Hate Speech, Insults, Misconduct
Extract Entities from Text	Sexual Content, Hate Speech, Insults, Misconduct
Classify Text into Categories	Sexual Content, Hate Speech, Insults, Misconduct
Generate and Match Text Metadata	Sexual Content, Hate Speech, Insults, Misconduct
Generate Text Summary	Sexual Content, Hate Speech, Insults, Misconduct
Generate Image Embeddings	❌ Not Covered
Generate Text Embeddings from Documents	❌ Not Covered

Prompt Injection Guardrails

In addition to content-safety guardrails, Knowledge Enrichment runs prompt injection guardrails on select actions. These guardrails inspect user-supplied inputs, such as instructions and other free-text parameters, for attempts to override or manipulate the underlying AI prompt. If an injection attempt is detected, the guardrails block the request.

The following table lists the actions with active prompt injection guardrails and the specific input fields (injection vectors) they monitor:

Task Type	Injection Vectors
Summarization	instructions
Generate Image Description	instructions
Extract Entities from Text	instructions
Detect Entities in Images	instructions
Classify Text into Categories	instructions, classes
Generate and Match Text Metadata	instructions, kSimilarMetadata
Classify Image into Categories	instructions, classes
Generate and Match Image Metadata	instructions, kSimilarMetadata

Impact of a Blocked Request

If your request triggers a guardrail, you may receive an error message like the one below. This means the system detected potentially unsafe content and stopped processing the request.

Guardrail Error Types

When a guardrail violation occurs during processing, the Context API returns an error in the action-specific result field. The error indicates that content triggered safety controls and processing was blocked.

GuardrailBlockedException

The GuardrailBlockedException error occurs when input or output content violates configured safety guardrails. This is a processing-level exception that prevents the action from completing.

Common causes:

• Input document contains content that violates guardrails (sexual content, hate speech, insults, misconduct).
• Generated output would violate guardrails.
• Instruction text includes prohibited content.

Resolution:

• Review the input document and remove or rephrase problematic content.
• Modify instruction text to avoid triggering guardrails.
• Contact support if the content appears to be blocked incorrectly.

---

Example: Guardrail Violation Error Response

The following is an example of an error response when a request is blocked due to guardrail violations:

{
  "id": "result_id",
  "timestamp": "2026-05-25T14:00:00Z",
  "results": [
    {
      "objectKey": "documents/sample.pdf",
      "imageDescription": null,
      "imageMetadata": null,
      "textMetadata": null,
      "textSummary": {
        "isSuccess": false,
        "result": null,
        "error": {
          "errorType": "GuardrailViolation",
          "message": "Processing for task text:summarization failed: ProcessingError (GuardrailBlockedException): This request triggered our default guardrails. To continue, please rephrase and submit again, or reach out to our support team for assistance."
        }
      },
      "textClassification": null,
      "imageClassification": null,
      "textEmbeddings": null,
      "imageEmbeddings": null,
      "generalProcessingErrors": null,
      "namedEntityText": null,
      "namedEntityImage": null
    }
  ],
  "status": "FAILURE",
  "inProgress": false
}