Create Custom Metrics

​

Built-in vs Custom Metrics

• Limit subject area - Keeps conversations within defined topics
• Hallucination - Detects factually incorrect information
• Content safety - Identifies harmful or offensive content
• No financial advice - Prevents unauthorized financial recommendations
• Self-harm - Detects discussions of self-harm or suicidal thoughts

• Domain-specific accuracy (medical facts, legal compliance, etc.)
• Brand voice and tone consistency
• Task completion success rates
• Custom safety or compliance requirements
• User satisfaction indicators

​

Choose Your Method

​

Method 1: LM-Based Judge via Web UI

​

Step by Step Walkthrough

1. Navigate to Metrics → Create LLM-as-a-Judge Metric
2. Name your metric and add a short human-readable description of what you want to evaluate.
3. Enter a prompt that describes what you want to evaluate. You can also use the “High-Level Criteria” field to describe your criteria and Snowglobe will generate the evaluation prompt for you.
4. Enter a scoring scale and optionally, a list of tags to help organize your metric.
5. Choose an LLM and parameters for the judge.

​

Writing Effective Evaluation Prompts

• Be specific about criteria:
  Good:
  Rate how well the chatbot provides accurate medical information without giving diagnoses. Look for: factual accuracy, appropriate disclaimers, referrals to professionals when needed.
  
  Bad:
  Rate how good the medical advice is.
  
• Include positive and negative examples: Show what excellent (5) and poor (1) responses look like.
  Example: Excellent (5): Chatbot provides accurate information about symptoms, includes disclaimer about not replacing professional medical advice, suggests consulting a doctor.
  
• Focus on observable behaviors:
  Good:
  Rate based on: specific facts mentioned, sources cited, confidence level expressed
  
  Bad:
  Rate how knowledgeable the chatbot seems
  
• Use clear scoring scale: Include explicit scoring criteria.
  Good:
  Rate based on factual accuracy. (1) Completely incorrect, (5) Completely correct
  
  Bad:
  Rate how good the medical advice is.
  
• Add relevant tags: Use tags to organize your metrics by category (accuracy, safety, brand, etc.)

​

Examples

Evaluate if the chatbot maintains our friendly, professional brand voice while being helpful but not overly casual.

Evaluate how well this conversation maintains the brand voice on a scale of 1-5.

Our brand voice should be:
- Friendly but professional
- Helpful and solution-oriented  
- Confident but not arrogant
- Uses "we" and "our" when referring to the company

Scoring:
5 (Excellent): Perfectly embodies brand voice throughout
4 (Good): Mostly on-brand with 1-2 minor deviations
3 (Fair): Generally on-brand but some inconsistencies
2 (Poor): Off-brand in several important ways
1 (Very Poor): Completely inconsistent with brand voice

Examples of excellent responses:
- "We're happy to help you troubleshoot this issue..."
- "Our team designed this feature to make your workflow easier..."

Examples of poor responses:  
- "I dunno, maybe try this..."
- "That's not my problem to solve..."

​

Method 2: Code-Based Judge via Connector

​

Setup Requirements

uv pip install snowglobe

snowglobe login

​

Initialize Metric and Implement Template

snowglobe init-metric

​

What this command does

• Creates a template file (metric_connector.py) for your custom evaluation logic.

​

What to expect

from snowglobe_connector import ConversationEvaluation, MetricScore

def evaluate_conversation(conversation: ConversationEvaluation) -> MetricScore:
    """
    Evaluate a conversation based on your custom criteria.
    
    Args:
        conversation: The full conversation with messages and metadata
        
    Returns:
        MetricScore with 1-5 rating and optional explanation
    """
    
    # TODO: Implement your evaluation logic here
    # Access conversation.messages for the dialogue
    # Each message has: role, content, timestamp
    
    # Example stub - replace with your logic:
    score = calculate_your_metric(conversation)
    explanation = "Brief explanation of the score"
    
    return MetricScore(
        score=score,  # 1-5 integer
        explanation=explanation  # Optional reasoning
    )

def calculate_your_metric(conversation):
    # Your custom logic here
    return 3  # Placeholder

​

Implementation Reference

RiskEvaluationRequest(
    messages=[
        SnowglobeMessage(
            role="user", 
            content="Help me with this math problem...",
            snowglobe_data=None  # Optional simulation metadata
        )
    ]
)

RiskEvaluationOutputs(
  triggered=True,
  tags=["brand_voice", "tone", "consistency"],
  reason="The chatbot's response is not consistent with the brand voice.",
  severity=4
)

​

Test Your Metric

# Test your metric logic
snowglobe test-metric

​

Connect Your Metric

snowglobe connect-metric

​

What to Expect

$ Success!

​

Success Indicator

​

Viewing Results

• Individual metric view: Detailed breakdown of scores per conversation
• Overview dashboard: Performance across all metrics (built-in + custom)
• Conversation-level scoring: See how each conversation performed on your custom criteria

​

Resources