Function TestRunBatchSimple

pkg/evals/batch_test.go:10–84

TestRunBatchSimple 测试简单批量评估

(t *testing.T)

// TestRunBatchSimple 测试简单批量评估
func TestRunBatchSimple(t *testing.T) {
	mockProvider := &MockProvider{
		response: `{"score": 0.85, "reason": "测试评分"}`,
	}

	testCases := []*BatchTestCase{
		{
			ID: "case1",
			Input: &TextEvalInput{
				Answer:  "巴黎是法国的首都。",
				Context: []string{"法国的首都是巴黎。"},
			},
		},
		{
			ID: "case2",
			Input: &TextEvalInput{
				Answer:  "伦敦是英国的首都。",
				Context: []string{"英国的首都是伦敦。"},
			},
		},
	}

	scorers := []Scorer{
		NewKeywordCoverageScorer(KeywordCoverageConfig{
			Keywords:        []string{"首都"},
			CaseInsensitive: true,
		}),
		NewLexicalSimilarityScorer(LexicalSimilarityConfig{
			MinTokenLength: 2,
		}),
		NewFaithfulnessScorer(mockProvider),
	}

	result, err := RunBatchSimple(context.Background(), testCases, scorers)
	if err != nil {
		t.Fatalf("RunBatchSimple() error = %v", err)
	}

	// 验证结果数量
	if len(result.Results) != 2 {
		t.Errorf("RunBatchSimple() got %d results, want 2", len(result.Results))
	}

	// 验证每个测试用例有 3 个 scorer 的结果
	for i, r := range result.Results {
		if len(r.Scores) != 3 {
			t.Errorf("Result[%d] got %d scores, want 3", i, len(r.Scores))
		}
		if r.Error != "" {
			t.Errorf("Result[%d] has error: %s", i, r.Error)
		}
	}

	// 验证汇总
	if result.Summary == nil {
		t.Fatal("RunBatchSimple() summary is nil")
	}