fix(knowledge): close root span on terminal state, enrich stage metadata

The root span created by OpenAttempt was never closed: PostProcess only
ended the postprocess stage, so the root row stayed at status=running
forever even after parse_status flipped to completed/failed. The
timeline rendered "进行中" indefinitely on the root, defeating the
whole "is the document done" question the timeline is meant to answer.

- SpanTracker.FinalizeAttempt(kid, attempt, status, output, code, msg):
  closes the root row idempotently. Re-closing a terminal root no-ops
  so success / cascade-fail / dead-letter paths can fire without
  coordination.
- PostProcess.Handle calls FinalizeAttempt(done) after EndSpan(postprocess)
  on the success path. Async downstream work (summary/question/wiki/
  graph) still records its own spans; their completion extends the
  trace's wall-clock end-time but doesn't reopen the root.
- FailSpan auto-closes the root when a MAIN pipeline stage fails
  (docreader / chunking / embedding / multimodal / postprocess).
  Cascade-cancelled siblings stay closed-with-the-cascade as before.
- Dead-letter callback (router/task.go) accepts the SpanTracker via
  DI and calls FinalizeAttempt(failed, TASK_TIMEOUT) when a
  document-related task exhausts retries. The probe payload now
  extracts the Attempt field that Document/Manual/PostProcess
  payloads already thread through.

Stage spans were also being recorded with nil input/output, leaving
the new detail panel with timestamps only. Each Begin/End site now
emits useful work metrics:
- DocReader input: file_name, file_type, is_url, url
              output: text_length, images_found, is_audio, pages
- Chunking input: chunks_planned
            output: chunks_written, total_text_chars
- Embedding input: chunks_to_embed, model_id, dim
             output: vectors_written, storage_bytes
- Multimodal input: image_count, enable_ocr, enable_caption
- PostProcess output: chunks_total, enqueued_summary, _question, _wiki, _graph

i18n: add knowledgeStages.root ("Knowledge processing") so the UI
can render a localized name instead of the raw span identifier.
This commit is contained in:
wizardchen
2026-05-27 16:54:09 +08:00
committed by lyingbug
parent c5722234ef
commit 4db6e69b4e
8 changed files with 218 additions and 20 deletions
+1
View File
@@ -390,6 +390,7 @@ export default {
knowledgeStages: {
title: 'Processing pipeline',
root: 'Knowledge processing',
attempt: 'Attempt {n}',
attemptLatest: 'Attempt {n} (latest)',
retry: 'Retry parsing',
+1
View File
@@ -396,6 +396,7 @@ export default {
},
knowledgeStages: {
title: "처리 파이프라인",
root: "지식 처리",
attempt: "{n}번째 시도",
attemptLatest: "{n}번째 시도 (최신)",
retry: "다시 파싱",
+1
View File
@@ -360,6 +360,7 @@ export default {
},
knowledgeStages: {
title: 'Конвейер обработки',
root: 'Обработка знаний',
attempt: 'Попытка {n}',
attemptLatest: 'Попытка {n} (последняя)',
retry: 'Повторить парсинг',
+1
View File
@@ -393,6 +393,7 @@ export default {
},
knowledgeStages: {
title: "处理流水线",
root: "知识处理",
attempt: "第 {n} 次尝试",
attemptLatest: "第 {n} 次尝试(最新)",
retry: "重新解析",
@@ -79,7 +79,10 @@ func (s *KnowledgePostProcessService) Handle(ctx context.Context, task *asynq.Ta
// and we never see the per-image fan-in here other than by reaching
// post-process). If the parent skipped multimodal entirely, the
// stage row will already be in "skipped" state and EndSpan is a
// no-op for missing rows.
// no-op for missing rows. Per-image success/failure counts are NOT
// aggregated here — the frontend already walks the children when
// rendering the multimodal stage detail and counts them itself,
// avoiding an extra query path.
if mm := s.tracker().LookupStage(ctx, payload.KnowledgeID, attempt, types.StageMultimodal); mm != nil &&
mm.Kind == types.SpanKindStage {
s.tracker().EndSpan(ctx, mm, nil)
@@ -137,16 +140,20 @@ func (s *KnowledgePostProcessService) Handle(ctx context.Context, task *asynq.Ta
}
// 4. Spawn Summary and Question Tasks
enqueuedSummary := false
enqueuedQuestion := false
if len(textChunks) > 0 {
s.enqueueSummaryGenerationTask(ctx, payload)
enqueuedSummary = true
// Question generation only makes sense for RAG indexing (improves chunk recall).
// Skip when only Wiki/Graph is enabled without vector/keyword search.
if kb.NeedsEmbeddingModel() {
s.enqueueQuestionGenerationIfEnabled(ctx, payload, kb)
enqueuedQuestion = s.enqueueQuestionGenerationIfEnabled(ctx, payload, kb)
}
}
// 5. Spawn Graph RAG Tasks — only when graph indexing is enabled in IndexingStrategy
enqueuedGraph := false
if kb.IsGraphEnabled() {
logger.Infof(ctx, "[KnowledgePostProcess] Spawning Graph RAG extract tasks for %d text-like chunks", len(textChunks))
for _, chunk := range textChunks {
@@ -155,16 +162,31 @@ func (s *KnowledgePostProcessService) Handle(ctx context.Context, task *asynq.Ta
logger.Errorf(ctx, "[KnowledgePostProcess] Failed to create chunk extract task for %s: %v", chunk.ID, err)
}
}
enqueuedGraph = len(textChunks) > 0
}
// 6. Spawn Wiki Ingest Task if wiki indexing is enabled in IndexingStrategy
enqueuedWiki := false
if kb.IndexingStrategy.WikiEnabled && len(textChunks) > 0 {
EnqueueWikiIngest(ctx, s.taskEnqueuer, s.pendingRepo, payload.TenantID, payload.KnowledgeBaseID, payload.KnowledgeID)
logger.Infof(ctx, "[KnowledgePostProcess] Enqueued wiki ingest task for %s", payload.KnowledgeID)
enqueuedWiki = true
}
s.tracker().EndSpan(ctx, postSpan, types.JSONMap{
"chunks_total": len(textChunks),
})
postOutput := types.JSONMap{
"chunks_total": len(textChunks),
"enqueued_summary": enqueuedSummary,
"enqueued_question": enqueuedQuestion,
"enqueued_wiki": enqueuedWiki,
"enqueued_graph": enqueuedGraph,
}
s.tracker().EndSpan(ctx, postSpan, postOutput)
// Close the root span — the parse pipeline is done. Async
// downstream stages (summary/question/wiki/graph) record their
// own spans independently; their finishing extends the trace's
// end-time but does not reopen the root. A late failure in one
// of those stages does not poison the parse result.
s.tracker().FinalizeAttempt(ctx, payload.KnowledgeID, attempt,
types.SpanStatusDone, postOutput, "", "")
return nil
}
@@ -194,13 +216,13 @@ func (s *KnowledgePostProcessService) enqueueSummaryGenerationTask(ctx context.C
}
}
func (s *KnowledgePostProcessService) enqueueQuestionGenerationIfEnabled(ctx context.Context, payload types.KnowledgePostProcessPayload, kb *types.KnowledgeBase) {
func (s *KnowledgePostProcessService) enqueueQuestionGenerationIfEnabled(ctx context.Context, payload types.KnowledgePostProcessPayload, kb *types.KnowledgeBase) bool {
if s.taskEnqueuer == nil {
return
return false
}
if kb.QuestionGenerationConfig == nil || !kb.QuestionGenerationConfig.Enabled {
return
return false
}
questionCount := kb.QuestionGenerationConfig.QuestionCount
@@ -222,13 +244,14 @@ func (s *KnowledgePostProcessService) enqueueQuestionGenerationIfEnabled(ctx con
payloadBytes, err := json.Marshal(taskPayload)
if err != nil {
logger.Warnf(ctx, "[KnowledgePostProcess] Failed to marshal question generation payload: %v", err)
return
return false
}
task := asynq.NewTask(types.TypeQuestionGeneration, payloadBytes, asynq.Queue("low"), asynq.MaxRetry(3))
if _, err := s.taskEnqueuer.Enqueue(task); err != nil {
logger.Warnf(ctx, "[KnowledgePostProcess] Failed to enqueue question generation for %s: %v", payload.KnowledgeID, err)
} else {
logger.Infof(ctx, "[KnowledgePostProcess] Enqueued question generation task for %s (count=%d)", payload.KnowledgeID, questionCount)
return false
}
logger.Infof(ctx, "[KnowledgePostProcess] Enqueued question generation task for %s (count=%d)", payload.KnowledgeID, questionCount)
return true
}
@@ -466,7 +466,9 @@ func (s *knowledgeService) processChunks(ctx context.Context,
// Chunks are needed for wiki generation, graph extraction, and summary generation
// even when vector/keyword indexing is disabled.
span.AddEvent("create chunks")
s.beginStage(ctx, knowledge.ID, types.StageChunking, nil)
s.beginStage(ctx, knowledge.ID, types.StageChunking, types.JSONMap{
"chunks_planned": len(insertChunks),
})
if err := s.chunkService.CreateChunks(ctx, insertChunks); err != nil {
knowledge.ParseStatus = types.ParseStatusFailed
knowledge.ErrorMessage = err.Error()
@@ -477,13 +479,27 @@ func (s *knowledgeService) processChunks(ctx context.Context,
werrors.ErrCodeChunkingFailed, "create chunks failed", err)
return
}
s.endStage(ctx, knowledge.ID, types.StageChunking, nil)
totalChunkChars := 0
for _, c := range insertChunks {
totalChunkChars += len(c.Content)
}
s.endStage(ctx, knowledge.ID, types.StageChunking, types.JSONMap{
"chunks_written": len(insertChunks),
"total_text_chars": totalChunkChars,
})
// Create index information and perform vector indexing — only when vector/keyword is enabled.
// Chunks are ALWAYS saved to DB (above) because wiki and graph need them even without vector indexing.
var totalStorageSize int64
if kb.NeedsEmbeddingModel() && embeddingModel != nil {
s.beginStage(ctx, knowledge.ID, types.StageEmbedding, nil)
embedInput := types.JSONMap{
"chunks_to_embed": len(textChunks),
"model_id": kb.EmbeddingModelID,
}
if dim := embeddingModel.GetDimensions(); dim > 0 {
embedInput["dim"] = dim
}
s.beginStage(ctx, knowledge.ID, types.StageEmbedding, embedInput)
// Create index information — only for child/flat chunks, NOT parent chunks.
// Parent chunks are stored for context retrieval but do not need vector embeddings.
// Prepend the document title to improve semantic alignment between
@@ -576,7 +592,10 @@ func (s *knowledgeService) processChunks(ctx context.Context,
return
}
logger.GetLogger(ctx).Infof("processChunks batch index successfully, with %d index", len(indexInfoList))
s.endStage(ctx, knowledge.ID, types.StageEmbedding, nil)
s.endStage(ctx, knowledge.ID, types.StageEmbedding, types.JSONMap{
"vectors_written": len(indexInfoList),
"storage_bytes": totalStorageSize,
})
// Final check before marking as completed - if deleted during processing, don't update status
if s.isKnowledgeDeleting(ctx, knowledge.TenantID, knowledge.ID) {
@@ -617,7 +636,11 @@ func (s *knowledgeService) processChunks(ctx context.Context,
// Enqueue multimodal tasks for images (async, non-blocking)
if options.EnableMultimodel && len(options.StoredImages) > 0 {
s.beginStage(ctx, knowledge.ID, types.StageMultimodal, nil)
s.beginStage(ctx, knowledge.ID, types.StageMultimodal, types.JSONMap{
"image_count": len(options.StoredImages),
"enable_ocr": true,
"enable_caption": true,
})
s.enqueueImageMultimodalTasks(ctx, knowledge, kb, options.StoredImages, chunks, options.Metadata)
} else {
s.skipStage(ctx, knowledge.ID, types.StageMultimodal, "skipped")
@@ -2336,7 +2359,15 @@ func (s *knowledgeService) convert(
// up — before that, the stage stays "pending" from the initial
// upload. Failure/skip transitions are emitted at the specific
// failure points below; success is emitted at the bottom.
s.beginStage(ctx, knowledge.ID, types.StageDocReader, nil)
docInput := types.JSONMap{
"file_name": payload.FileName,
"file_type": payload.FileType,
"is_url": payload.URL != "",
}
if payload.URL != "" {
docInput["url"] = payload.URL
}
s.beginStage(ctx, knowledge.ID, types.StageDocReader, docInput)
isURL := payload.URL != ""
fileType := payload.FileType
overrides := s.getParserEngineOverridesFromContext(ctx)
@@ -2426,7 +2457,15 @@ func (s *knowledgeService) convert(
werrors.ErrCodeDocReaderParseFailed, result.Error, nil)
return nil, nil
}
s.endStage(ctx, knowledge.ID, types.StageDocReader, nil)
docOutput := types.JSONMap{
"text_length": len(result.MarkdownContent),
"images_found": len(result.ImageRefs),
"is_audio": result.IsAudio,
}
if pages := result.Metadata["pages"]; pages != "" {
docOutput["pages"] = pages
}
s.endStage(ctx, knowledge.ID, types.StageDocReader, docOutput)
return result, nil
}
@@ -93,6 +93,15 @@ type SpanTracker interface {
// image_multimodal) attach subspans to the parent stage span
// created by the upstream pipeline.
LookupStage(ctx context.Context, knowledgeID string, attempt int, stage string) *Span
// FinalizeAttempt closes the root span for (knowledgeID, attempt)
// with the given terminal status (done | failed). Idempotent:
// re-closing an already-terminal root is a no-op so callers from
// multiple paths (success orchestrator, dead-letter handler,
// housekeeping) can fire without coordination. status defaults to
// done; output/error are written verbatim.
FinalizeAttempt(ctx context.Context, knowledgeID string, attempt int, status string,
output types.JSONMap, errorCode, errorMessage string)
}
type spanTracker struct {
@@ -456,6 +465,16 @@ func (t *spanTracker) FailSpan(ctx context.Context, span *Span, errorCode, error
// in StageDependencies (those are siblings, not descendants).
if span.Kind == types.SpanKindStage {
t.cascadeDependentStages(ctx, span, reason)
// Any failure in a MAIN pipeline stage means the attempt is
// done — the parse cannot succeed past this point. Close the
// root span as failed so the UI doesn't show "进行中" forever.
// Optional downstream stages (summary/question/wiki/graph) do
// not poison the attempt: they can fail without invalidating
// the parsed document.
if isMainPipelineStage(span.Name) {
t.FinalizeAttempt(ctx, span.KnowledgeID, span.Attempt,
types.SpanStatusFailed, nil, errorCode, errorMessage)
}
}
t.touchKnowledgeHeartbeat(ctx, span.KnowledgeID, span.Kind)
}
@@ -596,6 +615,21 @@ func contains(haystack []string, needle string) bool {
return false
}
// isMainPipelineStage reports whether stage is one of the 5 mandatory
// pipeline stages (docreader / chunking / embedding / multimodal /
// postprocess). A failure in any of these terminally invalidates the
// attempt and must close the root as failed. Optional downstream stages
// added later (summary, question, wiki, graph) do NOT match — those
// can fail individually without poisoning the parse result.
func isMainPipelineStage(name string) bool {
for _, s := range types.AllStages {
if s == name {
return true
}
}
return false
}
// durationSince computes elapsed ms preferring the in-process cache;
// falls back to the *Span's StartedAt for cross-process callers.
func durationSince(t *spanTracker, span *Span, now time.Time) int64 {
@@ -608,6 +642,88 @@ func durationSince(t *spanTracker, span *Span, now time.Time) int64 {
return 0
}
// FinalizeAttempt closes the root span for (knowledgeID, attempt). The
// pipeline calls this from two places: the success orchestrator
// (PostProcess) when the parse reaches Completed, and FailSpan when a
// MAIN stage fails terminally. Without this, the root row created by
// OpenAttempt would stay in "running" forever even after parse_status
// flips to completed/failed — operators would see a perpetually
// "running" trace despite a terminal knowledge state.
//
// Idempotent: re-closing a root that's already done/failed is a no-op
// so callers from different paths (success vs. cascade-fail vs.
// dead-letter) don't have to coordinate. We deliberately avoid the
// recordStart cache here (cross-process callers won't have it); we
// recompute duration from the persisted row's started_at.
func (t *spanTracker) FinalizeAttempt(ctx context.Context, knowledgeID string, attempt int, status string,
output types.JSONMap, errorCode, errorMessage string,
) {
if knowledgeID == "" || attempt <= 0 {
return
}
if status == "" {
status = types.SpanStatusDone
}
rows, err := t.repo.ListByAttempt(ctx, knowledgeID, attempt)
if err != nil {
logger.Warnf(ctx, "[SpanTracker] FinalizeAttempt list failed kid=%s attempt=%d: %v",
knowledgeID, attempt, err)
return
}
var root *types.KnowledgeProcessingSpan
for i := range rows {
if rows[i].Kind == types.SpanKindRoot {
cp := rows[i]
root = &cp
break
}
}
if root == nil {
// No root means nothing to close — likely an attempt that
// predates the tracker or whose OpenAttempt write failed.
return
}
if root.Status == types.SpanStatusDone || root.Status == types.SpanStatusFailed ||
root.Status == types.SpanStatusCancelled || root.Status == types.SpanStatusSkipped {
return
}
now := time.Now()
var started time.Time
if root.StartedAt != nil {
started = *root.StartedAt
}
dur := int64(0)
if !started.IsZero() {
dur = now.Sub(started).Milliseconds()
}
if len(errorMessage) > 1024 {
errorMessage = errorMessage[:1024]
}
row := &types.KnowledgeProcessingSpan{
KnowledgeID: root.KnowledgeID,
Attempt: root.Attempt,
SpanID: root.SpanID,
ParentSpanID: root.ParentSpanID,
Name: root.Name,
Kind: root.Kind,
Status: status,
Input: root.Input,
Output: output,
Metadata: root.Metadata,
ErrorCode: strings.TrimSpace(errorCode),
ErrorMessage: errorMessage,
StartedAt: root.StartedAt,
FinishedAt: &now,
DurationMs: dur,
}
if err := t.repo.Upsert(ctx, row); err != nil {
logger.Warnf(ctx, "[SpanTracker] FinalizeAttempt upsert failed kid=%s attempt=%d: %v",
knowledgeID, attempt, err)
return
}
t.touchKnowledgeHeartbeat(ctx, knowledgeID, types.SpanKindRoot)
}
// noopSpanTracker collapses every method to a no-op for tests/lite.
type noopSpanTracker struct{}
@@ -625,3 +741,5 @@ func (noopSpanTracker) EndSpan(_ context.Context, _ *Span, _ types.JSONMap)
func (noopSpanTracker) FailSpan(_ context.Context, _ *Span, _, _ string, _ error) {}
func (noopSpanTracker) SkipSpan(_ context.Context, _ *Span, _ string) {}
func (noopSpanTracker) LookupStage(_ context.Context, _ string, _ int, _ string) *Span { return nil }
func (noopSpanTracker) FinalizeAttempt(_ context.Context, _ string, _ int, _ string, _ types.JSONMap, _, _ string) {
}
+16 -2
View File
@@ -34,6 +34,7 @@ type AsynqTaskParams struct {
KnowledgePostProcess interfaces.TaskHandler `name:"knowledgePostProcess"`
WikiIngest interfaces.TaskHandler `name:"wikiIngest"`
DeadLetterRepo interfaces.TaskDeadLetterRepository
SpanTracker service.SpanTracker
}
// defaultRedisOpTimeout is the previous hard-coded read timeout. The 100ms
@@ -165,7 +166,7 @@ func RunAsynqServer(params AsynqTaskParams) *asynq.ServeMux {
// a permanently-failing task left its parent knowledge stranded in
// "processing" until housekeeping cron caught it minutes later — the
// UI signal users actually see.
knowledgeFailer := newDeadLetterKnowledgeFailer(params.KnowledgeService)
knowledgeFailer := newDeadLetterKnowledgeFailer(params.KnowledgeService, params.SpanTracker)
mux.Use(asynqdl.MiddlewareWithCallback(params.DeadLetterRepo, knowledgeFailer))
// Install Langfuse middleware BEFORE handler registration so every task
@@ -236,6 +237,12 @@ func RunAsynqServer(params AsynqTaskParams) *asynq.ServeMux {
// depend on the full payload schema and survive future field churn.
type deadLetterKnowledgePayload struct {
KnowledgeID string `json:"knowledge_id,omitempty"`
// Attempt threads through DocumentProcess / ManualProcess /
// KnowledgePostProcess payloads (added when span tracking shipped)
// — extracted here so the dead-letter callback can also close the
// matching root span as failed. Older in-flight payloads without
// this field decode as 0 and the tracker call no-ops.
Attempt int `json:"attempt,omitempty"`
}
// taskTypesAffectingKnowledgeStatus enumerates the asynq task types whose
@@ -269,7 +276,7 @@ var taskTypesAffectingKnowledgeStatus = map[string]struct{}{
// errors are logged and swallowed. The dead-letter record is the source of
// truth — this is purely a UX shortcut so users don't wait for the
// housekeeping cron's next sweep.
func newDeadLetterKnowledgeFailer(ks interfaces.KnowledgeService) asynqdl.OnDeadLetter {
func newDeadLetterKnowledgeFailer(ks interfaces.KnowledgeService, tracker service.SpanTracker) asynqdl.OnDeadLetter {
if ks == nil {
return nil
}
@@ -303,6 +310,13 @@ func newDeadLetterKnowledgeFailer(ks interfaces.KnowledgeService) asynqdl.OnDead
logger.Warnf(ctx, "dead-letter callback: failed to mark knowledge %s as failed: %v", probe.KnowledgeID, err)
return
}
// Close the matching root span so the timeline stops showing
// "进行中" after dead-letter exhaustion. Best-effort: nil
// tracker / missing attempt / missing root all no-op cleanly.
if tracker != nil && probe.Attempt > 0 {
tracker.FinalizeAttempt(ctx, probe.KnowledgeID, probe.Attempt,
types.SpanStatusFailed, nil, "TASK_TIMEOUT", errMsg)
}
logger.Infof(ctx, "dead-letter callback: marked knowledge %s as failed (task=%s)", probe.KnowledgeID, t.Type())
}
}