Function RunBatch

pkg/evals/batch.go:74–151

RunBatch 批量运行评估

(ctx context.Context, cfg *BatchConfig)

// RunBatch 批量运行评估
func RunBatch(ctx context.Context, cfg *BatchConfig) (*BatchEvalResult, error) {
	if len(cfg.TestCases) == 0 {
		return nil, errors.New("no test cases provided")
	}
	if len(cfg.Scorers) == 0 {
		return nil, errors.New("no scorers provided")
	}

	// 设置默认并发数
	if cfg.Concurrency <= 0 {
		cfg.Concurrency = 1
	}

	startTime := time.Now()
	results := make([]*BatchResult, len(cfg.TestCases))

	// 使用信号量控制并发
	sem := make(chan struct{}, cfg.Concurrency)
	var wg sync.WaitGroup
	var mu sync.Mutex
	completed := 0
	var firstErr error

	for i, testCase := range cfg.TestCases {
		// 检查是否应该停止
		mu.Lock()
		shouldStop := cfg.StopOnError && firstErr != nil
		mu.Unlock()
		if shouldStop {
			break
		}

		wg.Add(1)
		go func(index int, tc *BatchTestCase) {
			defer wg.Done()

			// 获取信号量
			sem <- struct{}{}
			defer func() { <-sem }()

			// 检查上下文是否已取消
			if ctx.Err() != nil {
				mu.Lock()
				if firstErr == nil {
					firstErr = ctx.Err()
				}
				mu.Unlock()
				return
			}

			// 运行单个测试用例
			result := runSingleTestCase(ctx, tc, cfg.Scorers)
			results[index] = result

			// 更新进度
			mu.Lock()
			completed++
			if result.Error != "" && firstErr == nil {