feat: 优化异步任务重试逻辑,仅在最后一次重试时更新失败状态,改进混合搜索结果融合策略

This commit is contained in:
wizardchen
2025-12-15 21:08:17 +08:00
committed by lyingbug
parent b20174efe2
commit efe7b5d0f1
2 changed files with 155 additions and 80 deletions
+67 -19
View File
@@ -4617,6 +4617,12 @@ func (s *knowledgeService) ProcessDocument(ctx context.Context, t *asynq.Task) e
ctx = logger.WithRequestID(ctx, payload.RequestId)
ctx = logger.WithField(ctx, "document_process", payload.KnowledgeID)
ctx = context.WithValue(ctx, types.TenantIDContextKey, payload.TenantID)
// 获取任务重试信息,用于判断是否是最后一次重试
retryCount, _ := asynq.GetRetryCount(ctx)
maxRetry, _ := asynq.GetMaxRetry(ctx)
isLastRetry := retryCount >= maxRetry
tenantInfo, err := s.tenantRepo.GetTenantByID(ctx, payload.TenantID)
if err != nil {
logger.Errorf(ctx, "failed to get tenant: %v", err)
@@ -4624,7 +4630,8 @@ func (s *knowledgeService) ProcessDocument(ctx context.Context, t *asynq.Task) e
}
ctx = context.WithValue(ctx, types.TenantInfoContextKey, tenantInfo)
logger.Infof(ctx, "Processing document task: knowledge_id=%s, file_path=%s", payload.KnowledgeID, payload.FilePath)
logger.Infof(ctx, "Processing document task: knowledge_id=%s, file_path=%s, retry=%d/%d",
payload.KnowledgeID, payload.FilePath, retryCount, maxRetry)
// 幂等性检查:获取knowledge记录
knowledge, err := s.repo.GetKnowledgeByID(ctx, payload.TenantID, payload.KnowledgeID)
@@ -4739,10 +4746,13 @@ func (s *knowledgeService) ProcessDocument(ctx context.Context, t *asynq.Task) e
RequestId: payload.RequestId,
})
if err != nil {
knowledge.ParseStatus = "failed"
knowledge.ErrorMessage = err.Error()
knowledge.UpdatedAt = time.Now()
s.repo.UpdateKnowledge(ctx, knowledge)
// 如果是最后一次重试,更新状态为失败
if isLastRetry {
knowledge.ParseStatus = "failed"
knowledge.ErrorMessage = err.Error()
knowledge.UpdatedAt = time.Now()
s.repo.UpdateKnowledge(ctx, knowledge)
}
return fmt.Errorf("failed to read from URL: %w", err)
}
chunks = urlResp.Chunks
@@ -4773,10 +4783,13 @@ func (s *knowledgeService) ProcessDocument(ctx context.Context, t *asynq.Task) e
if err != nil {
logger.GetLogger(ctx).WithField("knowledge_id", knowledge.ID).
WithField("error", err).Errorf("processDocument get file failed")
knowledge.ParseStatus = "failed"
knowledge.ErrorMessage = err.Error()
knowledge.UpdatedAt = time.Now()
s.repo.UpdateKnowledge(ctx, knowledge)
// 如果是最后一次重试,更新状态为失败
if isLastRetry {
knowledge.ParseStatus = "failed"
knowledge.ErrorMessage = err.Error()
knowledge.UpdatedAt = time.Now()
s.repo.UpdateKnowledge(ctx, knowledge)
}
return fmt.Errorf("failed to get file: %w", err)
}
defer fileReader.Close()
@@ -4784,10 +4797,13 @@ func (s *knowledgeService) ProcessDocument(ctx context.Context, t *asynq.Task) e
// 读取文件内容
contentBytes, err := io.ReadAll(fileReader)
if err != nil {
knowledge.ParseStatus = "failed"
knowledge.ErrorMessage = err.Error()
knowledge.UpdatedAt = time.Now()
s.repo.UpdateKnowledge(ctx, knowledge)
// 如果是最后一次重试,更新状态为失败
if isLastRetry {
knowledge.ParseStatus = "failed"
knowledge.ErrorMessage = err.Error()
knowledge.UpdatedAt = time.Now()
s.repo.UpdateKnowledge(ctx, knowledge)
}
return fmt.Errorf("failed to read file: %w", err)
}
@@ -4817,10 +4833,13 @@ func (s *knowledgeService) ProcessDocument(ctx context.Context, t *asynq.Task) e
if err != nil {
logger.GetLogger(ctx).WithField("knowledge_id", knowledge.ID).
WithField("error", err).Errorf("processDocument read file failed")
knowledge.ParseStatus = "failed"
knowledge.ErrorMessage = err.Error()
knowledge.UpdatedAt = time.Now()
s.repo.UpdateKnowledge(ctx, knowledge)
// 如果是最后一次重试,更新状态为失败
if isLastRetry {
knowledge.ParseStatus = "failed"
knowledge.ErrorMessage = err.Error()
knowledge.UpdatedAt = time.Now()
s.repo.UpdateKnowledge(ctx, knowledge)
}
return fmt.Errorf("failed to read file from docreader: %w", err)
}
chunks = fileResp.Chunks
@@ -4847,6 +4866,11 @@ func (s *knowledgeService) ProcessFAQImport(ctx context.Context, t *asynq.Task)
ctx = logger.WithField(ctx, "faq_import", payload.TaskID)
ctx = context.WithValue(ctx, types.TenantIDContextKey, payload.TenantID)
// 获取任务重试信息,用于判断是否是最后一次重试
retryCount, _ := asynq.GetRetryCount(ctx)
maxRetry, _ := asynq.GetMaxRetry(ctx)
isLastRetry := retryCount >= maxRetry
tenantInfo, err := s.tenantRepo.GetTenantByID(ctx, payload.TenantID)
if err != nil {
logger.Errorf(ctx, "failed to get tenant: %v", err)
@@ -4854,8 +4878,8 @@ func (s *knowledgeService) ProcessFAQImport(ctx context.Context, t *asynq.Task)
}
ctx = context.WithValue(ctx, types.TenantInfoContextKey, tenantInfo)
logger.Infof(ctx, "Processing FAQ import task: task_id=%s, kb_id=%s, total_entries=%d",
payload.TaskID, payload.KBID, len(payload.Entries))
logger.Infof(ctx, "Processing FAQ import task: task_id=%s, kb_id=%s, total_entries=%d, retry=%d/%d",
payload.TaskID, payload.KBID, len(payload.Entries), retryCount, maxRetry)
// 幂等性检查:获取knowledge记录(FAQ任务使用knowledge ID作为taskID)
knowledge, err := s.repo.GetKnowledgeByID(ctx, payload.TenantID, payload.TaskID)
@@ -4871,6 +4895,12 @@ func (s *knowledgeService) ProcessFAQImport(ctx context.Context, t *asynq.Task)
kb, err := s.kbService.GetKnowledgeBaseByID(ctx, payload.KBID)
if err != nil {
logger.Errorf(ctx, "Failed to get knowledge base: %v", err)
// 如果是最后一次重试,更新状态为失败
if isLastRetry {
if updateErr := s.updateFAQImportStatus(ctx, payload.TaskID, types.FAQImportStatusFailed, 0, len(payload.Entries), 0, err.Error()); updateErr != nil {
logger.Errorf(ctx, "Failed to update task status to failed: %v", updateErr)
}
}
return fmt.Errorf("failed to get knowledge base: %w", err)
}
@@ -4897,6 +4927,12 @@ func (s *knowledgeService) ProcessFAQImport(ctx context.Context, t *asynq.Task)
chunksDeleted, err := s.chunkRepo.DeleteUnindexedChunks(ctx, payload.TenantID, payload.KnowledgeID)
if err != nil {
logger.Errorf(ctx, "Failed to delete unindexed chunks: %v", err)
// 如果是最后一次重试,更新状态为失败
if isLastRetry {
if updateErr := s.updateFAQImportStatus(ctx, payload.TaskID, types.FAQImportStatusFailed, 0, originalTotalEntries, processedCount, err.Error()); updateErr != nil {
logger.Errorf(ctx, "Failed to update task status to failed: %v", updateErr)
}
}
return fmt.Errorf("failed to delete unindexed chunks: %w", err)
}
logger.Infof(ctx, "Deleted unindexed chunks: %d", len(chunksDeleted))
@@ -4949,6 +4985,18 @@ func (s *knowledgeService) ProcessFAQImport(ctx context.Context, t *asynq.Task)
// 执行FAQ导入
if err := s.executeFAQImport(ctx, payload.TaskID, payload.KBID, faqPayload, payload.TenantID, originalTotalEntries-len(payload.Entries)); err != nil {
logger.Errorf(ctx, "FAQ import task failed: %s, error: %v", payload.TaskID, err)
// 如果是最后一次重试,更新状态为失败
if isLastRetry {
// 获取当前已处理的进度
currentMeta, _ := types.ParseFAQImportMetadata(knowledge)
currentProcessed := 0
if currentMeta != nil {
currentProcessed = currentMeta.ImportProcessed
}
if updateErr := s.updateFAQImportStatus(ctx, payload.TaskID, types.FAQImportStatusFailed, 0, originalTotalEntries, currentProcessed, err.Error()); updateErr != nil {
logger.Errorf(ctx, "Failed to update task status to failed: %v", updateErr)
}
}
return fmt.Errorf("FAQ import failed: %w", err)
}
+88 -61
View File
@@ -544,81 +544,108 @@ func (s *knowledgeBaseService) HybridSearch(ctx context.Context,
logger.Info(ctx, "No search results found")
return nil, nil
}
logger.Infof(ctx, "Result count before RRF fusion: vector=%d, keyword=%d", len(vectorResults), len(keywordResults))
logger.Infof(ctx, "Result count before fusion: vector=%d, keyword=%d", len(vectorResults), len(keywordResults))
// Use RRF (Reciprocal Rank Fusion) to merge results
// RRF score = sum(1 / (k + rank)) for each retriever where the chunk appears
// k=60 is a common choice that works well in practice
const rrfK = 60
var deduplicatedChunks []*types.IndexWithScore
// Build rank maps for each retriever (already sorted by score from retriever)
vectorRanks := make(map[string]int)
for i, r := range vectorResults {
if _, exists := vectorRanks[r.ChunkID]; !exists {
vectorRanks[r.ChunkID] = i + 1 // 1-indexed rank
// If only vector results (no keyword results), keep original embedding scores
// This is important for FAQ search which only uses vector retrieval
if len(keywordResults) == 0 {
logger.Info(ctx, "Only vector results, keeping original embedding scores")
chunkInfoMap := make(map[string]*types.IndexWithScore)
for _, r := range vectorResults {
if _, exists := chunkInfoMap[r.ChunkID]; !exists {
chunkInfoMap[r.ChunkID] = r
}
}
}
keywordRanks := make(map[string]int)
for i, r := range keywordResults {
if _, exists := keywordRanks[r.ChunkID]; !exists {
keywordRanks[r.ChunkID] = i + 1 // 1-indexed rank
deduplicatedChunks = make([]*types.IndexWithScore, 0, len(chunkInfoMap))
for _, info := range chunkInfoMap {
deduplicatedChunks = append(deduplicatedChunks, info)
}
}
slices.SortFunc(deduplicatedChunks, func(a, b *types.IndexWithScore) int {
if a.Score > b.Score {
return -1
} else if a.Score < b.Score {
return 1
}
return 0
})
logger.Infof(ctx, "Result count after deduplication: %d", len(deduplicatedChunks))
} else {
// Use RRF (Reciprocal Rank Fusion) to merge results from multiple retrievers
// RRF score = sum(1 / (k + rank)) for each retriever where the chunk appears
// k=60 is a common choice that works well in practice
const rrfK = 60
// Collect all unique chunks and compute RRF scores
chunkInfoMap := make(map[string]*types.IndexWithScore)
rrfScores := make(map[string]float64)
// Process vector results
for _, r := range vectorResults {
if _, exists := chunkInfoMap[r.ChunkID]; !exists {
chunkInfoMap[r.ChunkID] = r
// Build rank maps for each retriever (already sorted by score from retriever)
vectorRanks := make(map[string]int)
for i, r := range vectorResults {
if _, exists := vectorRanks[r.ChunkID]; !exists {
vectorRanks[r.ChunkID] = i + 1 // 1-indexed rank
}
}
}
// Process keyword results
for _, r := range keywordResults {
if _, exists := chunkInfoMap[r.ChunkID]; !exists {
chunkInfoMap[r.ChunkID] = r
keywordRanks := make(map[string]int)
for i, r := range keywordResults {
if _, exists := keywordRanks[r.ChunkID]; !exists {
keywordRanks[r.ChunkID] = i + 1 // 1-indexed rank
}
}
}
// Compute RRF scores
for chunkID := range chunkInfoMap {
rrfScore := 0.0
if rank, ok := vectorRanks[chunkID]; ok {
rrfScore += 1.0 / float64(rrfK+rank)
// Collect all unique chunks and compute RRF scores
chunkInfoMap := make(map[string]*types.IndexWithScore)
rrfScores := make(map[string]float64)
// Process vector results
for _, r := range vectorResults {
if _, exists := chunkInfoMap[r.ChunkID]; !exists {
chunkInfoMap[r.ChunkID] = r
}
}
if rank, ok := keywordRanks[chunkID]; ok {
rrfScore += 1.0 / float64(rrfK+rank)
// Process keyword results
for _, r := range keywordResults {
if _, exists := chunkInfoMap[r.ChunkID]; !exists {
chunkInfoMap[r.ChunkID] = r
}
}
rrfScores[chunkID] = rrfScore
}
// Convert to slice and sort by RRF score
deduplicatedChunks := make([]*types.IndexWithScore, 0, len(chunkInfoMap))
for chunkID, info := range chunkInfoMap {
// Store RRF score in the Score field for downstream processing
info.Score = rrfScores[chunkID]
deduplicatedChunks = append(deduplicatedChunks, info)
}
slices.SortFunc(deduplicatedChunks, func(a, b *types.IndexWithScore) int {
if a.Score > b.Score {
return -1
} else if a.Score < b.Score {
return 1
// Compute RRF scores
for chunkID := range chunkInfoMap {
rrfScore := 0.0
if rank, ok := vectorRanks[chunkID]; ok {
rrfScore += 1.0 / float64(rrfK+rank)
}
if rank, ok := keywordRanks[chunkID]; ok {
rrfScore += 1.0 / float64(rrfK+rank)
}
rrfScores[chunkID] = rrfScore
}
return 0
})
logger.Infof(ctx, "Result count after RRF fusion: %d", len(deduplicatedChunks))
// Convert to slice and sort by RRF score
deduplicatedChunks = make([]*types.IndexWithScore, 0, len(chunkInfoMap))
for chunkID, info := range chunkInfoMap {
// Store RRF score in the Score field for downstream processing
info.Score = rrfScores[chunkID]
deduplicatedChunks = append(deduplicatedChunks, info)
}
slices.SortFunc(deduplicatedChunks, func(a, b *types.IndexWithScore) int {
if a.Score > b.Score {
return -1
} else if a.Score < b.Score {
return 1
}
return 0
})
// Log top results after RRF fusion for debugging
for i, chunk := range deduplicatedChunks {
if i < 15 {
vRank, vOk := vectorRanks[chunk.ChunkID]
kRank, kOk := keywordRanks[chunk.ChunkID]
logger.Debugf(ctx, "RRF rank %d: chunk_id=%s, rrf_score=%.6f, vector_rank=%v(%v), keyword_rank=%v(%v)",
i, chunk.ChunkID, chunk.Score, vRank, vOk, kRank, kOk)
logger.Infof(ctx, "Result count after RRF fusion: %d", len(deduplicatedChunks))
// Log top results after RRF fusion for debugging
for i, chunk := range deduplicatedChunks {
if i < 15 {
vRank, vOk := vectorRanks[chunk.ChunkID]
kRank, kOk := keywordRanks[chunk.ChunkID]
logger.Debugf(ctx, "RRF rank %d: chunk_id=%s, rrf_score=%.6f, vector_rank=%v(%v), keyword_rank=%v(%v)",
i, chunk.ChunkID, chunk.Score, vRank, vOk, kRank, kOk)
}
}
}