evaluation

type BenchmarkResult struct {
	TotalCases      int           `json:"total_cases"`
	AvgFaithfulness float32       `json:"avg_faithfulness"`
	AvgRelevance    float32       `json:"avg_relevance"`
	AvgPrecision    float32       `json:"avg_precision"`
	TotalDuration   time.Duration `json:"total_duration"`
	Results         []CaseResult  `json:"results"`
}

type CaseResult struct {
	Query             string        `json:"query"`
	Answer            string        `json:"answer"`
	FaithfulnessScore float32       `json:"faithfulness"`
	RelevanceScore    float32       `json:"relevance"`
	PrecisionScore    float32       `json:"precision"`
	Duration          time.Duration `json:"duration"`
}

type LLMJudge interface {
	// EvaluateFaithfulness checks if the generated answer is strictly grounded in the retrieved chunks.
	EvaluateFaithfulness(ctx context.Context, query string, chunks []*core.Chunk, answer string) (score float32, reason string, err error)

	// EvaluateAnswerRelevance checks if the answer effectively addresses the user's intent.
	EvaluateAnswerRelevance(ctx context.Context, query string, answer string) (score float32, reason string, err error)

	// EvaluateContextPrecision checks if the retrieved context actually contains the useful information.
	EvaluateContextPrecision(ctx context.Context, query string, chunks []*core.Chunk) (score float32, reason string, err error)
}

type Label = core.CRAGLabel

type RagasLLMJudge struct {
	// contains filtered or unexported fields
}

type TestCase struct {
	Query       string `json:"query"`
	GroundTruth string `json:"ground_truth,omitempty"` // Optional: What we expect the answer to be
}