Function TestCalculateSummary

pkg/evals/batch_test.go:279–335

TestCalculateSummary 测试汇总计算

(t *testing.T)

// TestCalculateSummary 测试汇总计算
func TestCalculateSummary(t *testing.T) {
	results := []*BatchResult{
		{
			TestCaseID: "case1",
			Scores: []*ScoreResult{
				{Name: "scorer1", Value: 0.8},
				{Name: "scorer2", Value: 0.9},
			},
			Duration: 100 * time.Millisecond,
			Error:    "",
		},
		{
			TestCaseID: "case2",
			Scores: []*ScoreResult{
				{Name: "scorer1", Value: 0.7},
				{Name: "scorer2", Value: 0.85},
			},
			Duration: 150 * time.Millisecond,
			Error:    "",
		},
		{
			TestCaseID: "case3",
			Scores:     []*ScoreResult{},
			Duration:   0,
			Error:      "test error",
		},
	}

	summary := calculateSummary(results)

	if summary.TotalCases != 3 {
		t.Errorf("TotalCases = %d, want 3", summary.TotalCases)
	}

	if summary.SuccessfulCases != 2 {
		t.Errorf("SuccessfulCases = %d, want 2", summary.SuccessfulCases)
	}

	if summary.FailedCases != 1 {
		t.Errorf("FailedCases = %d, want 1", summary.FailedCases)
	}

	// 验证平均分
	if avgScore1, ok := summary.AverageScores["scorer1"]; !ok || avgScore1 != 0.75 {
		t.Errorf("AverageScores[scorer1] = %v, want 0.75", avgScore1)
	}

	if avgScore2, ok := summary.AverageScores["scorer2"]; !ok || avgScore2 != 0.875 {
		t.Errorf("AverageScores[scorer2] = %v, want 0.875", avgScore2)
	}

	// 验证平均执行时间
	expectedAvgDuration := 125 * time.Millisecond
	if summary.AverageDuration != expectedAvgDuration {
		t.Errorf("AverageDuration = %v, want %v", summary.AverageDuration, expectedAvgDuration)
	}
}