AgentEval metrics.instructions.md

Metric Implementation Guidelines

Metric Naming Prefixes

ALL metrics MUST use these prefixes in their Name property:

• llm_ = LLM-evaluated (costs API calls) → llm_faithfulness
• code_ = Computed by code (free) → code_tool_success
• embed_ = Embedding-based (costs embedding API) → embed_answer_similarity

Interface Selection

• IRAGMetric - For metrics that evaluate retrieval-augmented generation
  • Set RequiresContext = true if context documents needed
  • Set RequiresGroundTruth = true if expected answer needed
• IAgenticMetric - For metrics that evaluate agent tool usage
  • Set RequiresToolUsage = true

Standard Metric Structure

public class MyMetric : IRAGMetric
{
    private readonly IChatClient _evaluator;
    
    public string Name => "llm_my_metric";
    public string Description => "Evaluates XYZ quality";
    public bool RequiresContext => true;
    public bool RequiresGroundTruth => false;
    
    public MyMetric(IChatClient evaluator) => _evaluator = evaluator;
    
    public async Task<MetricResult> EvaluateAsync(
        EvaluationContext context, 
        CancellationToken ct = default)
    {
        // Validate required fields
        if (string.IsNullOrEmpty(context.Context))
            return MetricResult.Fail(Name, "Context is required");
        
        // Call LLM for evaluation
        var prompt = BuildEvaluationPrompt(context);
        var response = await _evaluator.GetResponseAsync([...], ct: ct);
        
        // Parse response (use LlmJsonParser for JSON responses)
        var parsed = LlmJsonParser.Parse<EvalResponse>(response.Text);
        
        return MetricResult.Pass(Name, parsed.Score, parsed.Explanation);
    }
}

Using LlmJsonParser

For LLM responses that return JSON, use the built-in parser:

var result = LlmJsonParser.Parse<T>(responseText);
// Handles markdown code blocks, extracts JSON, deserializes

Score Normalization

All scores should be 0-100 scale. Use ScoreNormalizer:

var normalized = ScoreNormalizer.From1To5(rawScore); // 1-5 → 0-100
var normalized = ScoreNormalizer.From0To1(rawScore); // 0-1 → 0-100