Skip to content

Message Assessment API

The Message Assessment API evaluates assistant messages for quality, hallucinations, and compliance.

Overview

Create assessments to track:

  • Hallucination detection
  • Compliance violations
  • Response quality
  • Factual accuracy

Methods

unique_sdk.MessageAssessment.create - Create assessment

Create a new assessment for an assistant message.

Parameters:

  • messageId (str, required) - Message ID to assess
  • status (Literal["PENDING", "DONE", "ERROR"], required) - Assessment status
  • type (Literal["HALLUCINATION", "COMPLIANCE"], required) - Assessment type
  • isVisible (bool, required) - Whether to show assessment to user
  • title (str | None, optional) - Assessment title
  • explanation (str | None, optional) - Detailed explanation
  • label (Literal["RED", "YELLOW", "GREEN"] | None, optional) - Color label

Returns:

Returns a MessageAssessment object.

Example:

assessment = unique_sdk.MessageAssessment.create(
    user_id=user_id,
    company_id=company_id,
    assistant_message_id="msg_abc123",
    status="DONE",
    label="RED",
    type="HALLUCINATION",
    title="Hallucination detected",
    explanation="The response contains information not found in source documents.",
    isVisible=True
)
unique_sdk.MessageAssessment.modify - Update assessment

Update an existing assessment.

Parameters:

  • messageId (str, required) - Message ID of the assessment to update
  • type (Literal["HALLUCINATION", "COMPLIANCE"], required) - Assessment type
  • status (Literal["PENDING", "DONE", "ERROR"] | None, optional) - Updated status
  • title (str | None, optional) - Updated title
  • explanation (str | None, optional) - Updated explanation
  • label (Literal["RED", "YELLOW", "GREEN"] | None, optional) - Updated label

Returns:

Returns a MessageAssessment object.

Example:

assessment = unique_sdk.MessageAssessment.modify(
    user_id=user_id,
    company_id=company_id,
    assistant_message_id="msg_abc123",
    status="DONE",
    label="YELLOW",
    type="HALLUCINATION",
    title="Minor inconsistency found",
    explanation="Updated assessment after review."
)

Use Cases

Hallucination Detection 
def detect_hallucination(assistant_message_id, response_text, source_documents):
    """Check for hallucinations in assistant response."""

    # Analyze response against sources
    has_hallucination, details = check_against_sources(
        response_text,
        source_documents
    )

    if has_hallucination:
        # Create assessment
        unique_sdk.MessageAssessment.create(
            user_id=user_id,
            company_id=company_id,
            messageId=assistant_message_id,
            status="DONE",
            label="RED",
            type="HALLUCINATION",
            title="Hallucination Detected",
            explanation=f"Found unsupported claims: {details}",
            isVisible=True
        )
    else:
        # Create positive assessment
        unique_sdk.MessageAssessment.create(
            user_id=user_id,
            company_id=company_id,
            messageId=assistant_message_id,
            status="DONE",
            label="GREEN",
            type="HALLUCINATION",
            title="No Hallucinations",
            explanation="All claims are supported by source documents.",
            isVisible=False  # Don't show positive checks to user
        )
Compliance Checking 
def check_compliance(assistant_message_id, response_text):
    """Check response for compliance violations."""

    violations = []

    # Check for PII
    if contains_pii(response_text):
        violations.append("Contains Personal Identifiable Information")

    # Check for prohibited content
    if contains_prohibited_content(response_text):
        violations.append("Contains prohibited content")

    # Check for policy violations
    if violates_policy(response_text):
        violations.append("Violates company policy")

    if violations:
        unique_sdk.MessageAssessment.create(
            user_id=user_id,
            company_id=company_id,
            messageId=assistant_message_id,
            status="DONE",
            label="RED",
            type="COMPLIANCE",
            title="Compliance Violations",
            explanation="; ".join(violations),
            isVisible=True
        )
        return False

    return True

Best Practices

Clear Explanations 
# Good: Specific and actionable
unique_sdk.MessageAssessment.create(
    ...
    title="Unsupported Claim Detected",
    explanation="The response states 'revenue increased by 40%' but source document shows 30%. See paragraph 3 in Q4_Report.pdf"
)

# Bad: Vague
unique_sdk.MessageAssessment.create(
    ...
    title="Error",
    explanation="Something is wrong"
)
Appropriate Visibility 
# Show critical issues to users
unique_sdk.MessageAssessment.create(
    label="RED",
    isVisible=True,  # User should see this
    ...
)

# Hide positive checks
unique_sdk.MessageAssessment.create(
    label="GREEN",
    isVisible=False,  # No need to show passing checks
    ...
)
Consistent Labeling 
# Define constants for consistency
class AssessmentLabel:
    CRITICAL = "RED"      # Serious issues
    WARNING = "YELLOW"    # Minor concerns
    PASSED = "GREEN"      # No issues

# Use consistently
unique_sdk.MessageAssessment.create(
    label=AssessmentLabel.CRITICAL,
    ...
)

Return Types

MessageAssessment

The MessageAssessment object represents an assessment of a message

Fields:

  • id (str) - Unique assessment identifier
  • messageId (str) - Associated message ID
  • status (Literal["PENDING", "DONE", "ERROR"]) - Assessment status
  • type (Literal["HALLUCINATION", "COMPLIANCE"]) - Assessment type
  • isVisible (bool) - Whether assessment is visible to user
  • title (str | None) - Assessment title
  • explanation (str | None) - Detailed explanation
  • label (Literal["RED", "YELLOW", "GREEN"] | None) - Color label
  • createdAt (str | None) - Creation timestamp (ISO 8601)
  • updatedAt (str | None) - Last update timestamp (ISO 8601)

Returned by: MessageAssessment.create(), MessageAssessment.modify()