Merge f7c0fe273b into 3032384457

优化摘要生成逻辑：
1. 将chunk token限制从16KB提升到30KB，更合理地利用模型上下文
2. 简化摘要聚合逻辑，移除复杂的分批聚合，直接合并所有summaries
3. 保留fallback机制，当最终摘要生成失败时返回已聚合的摘要

这些改进确保了：
- 长文档能够更充分地被摘要（30KB vs 16KB）
- 代码更简洁，避免不必要的迭代聚合和额外LLM调用
- 即使最终摘要失败也能返回有用的结果

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-authored-by: monkeycode-ai <monkeycode-ai@chaitin.com>
Co-authored-by: monkeycode-ai <monkeycode-ai@chaitin.com>

backend/usecase/llm.go

@@ -39,7 +39,6 @@ type LLMUsecase struct {
 const (
 	summaryChunkTokenLimit = 30720 // 30KB tokens per chunk
 	summaryMaxChunks       = 4     // max chunks to process for summary
-	summaryAggregateLimit  = 8192  // max tokens for aggregating summaries
 )
 
 func NewLLMUsecase(config *config.Config, rag rag.RAGService, conversationRepo *pg.ConversationRepository, kbRepo *pg.KnowledgeBaseRepository, nodeRepo *pg.NodeRepository, modelRepo *pg.ModelRepository, promptRepo *pg.PromptRepo, logger *log.Logger) *LLMUsecase {
@@ -227,39 +226,7 @@ func (u *LLMUsecase) SummaryNode(ctx context.Context, model *domain.Model, name,
 		return "", fmt.Errorf("failed to generate summary for document %s", name)
 	}
 
-	// Iteratively aggregate summaries if needed
-	for len(summaries) > 1 {
-		joined := strings.Join(summaries, "\n\n")
-		tokens, err := u.countTokens(joined)
-		if err != nil {
-			u.logger.Warn("Failed to count tokens for aggregation, proceeding anyway", log.Error(err))
-			break
-		}
-		if tokens <= summaryAggregateLimit {
-			break
-		}
-		// If still too large, aggregate in batches
-		u.logger.Debug("aggregating summaries in batches", log.Int("current_summaries", len(summaries)), log.Int("tokens", tokens))
-		batchSize := 2
-		newSummaries := make([]string, 0, (len(summaries)+batchSize-1)/batchSize)
-		for i := 0; i < len(summaries); i += batchSize {
-			end := i + batchSize
-			if end > len(summaries) {
-				end = len(summaries)
-			}
-			batch := strings.Join(summaries[i:end], "\n\n")
-			summary, err := u.requestSummary(ctx, chatModel, name, batch)
-			if err != nil {
-				u.logger.Error("Failed to aggregate summary batch", log.Int("batch_start", i), log.Error(err))
-				// Fallback: use the first summary in the batch
-				newSummaries = append(newSummaries, summaries[i])
-				continue
-			}
-			newSummaries = append(newSummaries, summary)
-		}
-		summaries = newSummaries
-	}
-
+	// Join all summaries and generate final summary
 	joined := strings.Join(summaries, "\n\n")
 	finalSummary, err := u.requestSummary(ctx, chatModel, name, joined)
 	if err != nil {
@@ -301,15 +268,6 @@ func (u *LLMUsecase) requestSummary(ctx context.Context, chatModel model.BaseCha
 	return strings.TrimSpace(u.trimThinking(summary)), nil
 }
 
-func (u *LLMUsecase) countTokens(text string) (int, error) {
-	encoding, err := tiktoken.GetEncoding("cl100k_base")
-	if err != nil {
-		return 0, fmt.Errorf("failed to get encoding: %w", err)
-	}
-	tokens := encoding.Encode(text, nil, nil)
-	return len(tokens), nil
-}
-
 func (u *LLMUsecase) SplitByTokenLimit(text string, maxTokens int) ([]string, error) {
 	if maxTokens <= 0 {
 		return nil, fmt.Errorf("maxTokens must be greater than 0")