Skip to content

Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.

29.8 KB
Raw Blame
package service

import (
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"sort"
	"time"

	"github.com/Tencent/WeKnora/internal/application/access"
	"github.com/Tencent/WeKnora/internal/infrastructure/docparser"
	"github.com/Tencent/WeKnora/internal/logger"
	"github.com/Tencent/WeKnora/internal/tracing/langfuse"
	"github.com/Tencent/WeKnora/internal/types"
	"github.com/Tencent/WeKnora/internal/types/interfaces"
	"github.com/hibiken/asynq"
	"github.com/redis/go-redis/v9"
)

// KnowledgePostProcessService acts as an orchestrator for all post-processing tasks
// after a document has been parsed and split into chunks (including multimodal OCR/Caption).
type KnowledgePostProcessService struct {
	knowledgeRepo interfaces.KnowledgeRepository
	kbService     interfaces.KnowledgeBaseService
	chunkService  interfaces.ChunkService
	chunkRepo     interfaces.ChunkRepository
	taskEnqueuer  interfaces.TaskEnqueuer
	pendingRepo   interfaces.TaskPendingOpsRepository
	redisClient   *redis.Client
	spanTracker   SpanTracker
}

func NewKnowledgePostProcessService(
	knowledgeRepo interfaces.KnowledgeRepository,
	kbService interfaces.KnowledgeBaseService,
	chunkService interfaces.ChunkService,
	chunkRepo interfaces.ChunkRepository,
	taskEnqueuer interfaces.TaskEnqueuer,
	pendingRepo interfaces.TaskPendingOpsRepository,
	redisClient *redis.Client,
	spanTracker SpanTracker,
) interfaces.TaskHandler {
	return &KnowledgePostProcessService{
		knowledgeRepo: knowledgeRepo,
		kbService:     kbService,
		chunkService:  chunkService,
		chunkRepo:     chunkRepo,
		taskEnqueuer:  taskEnqueuer,
		pendingRepo:   pendingRepo,
		redisClient:   redisClient,
		spanTracker:   spanTracker,
	}
}

func (s *KnowledgePostProcessService) tracker() SpanTracker {
	if s.spanTracker == nil {
		return noopSpanTracker{}
	}
	return s.spanTracker
}

// finishRunningMultimodalStage closes the multimodal stage only when image
// work really ran and is still open. The canonical stage also exists when
// multimodal processing is disabled, but that row is already "skipped" and
// must not be rewritten to "done" with the postprocess queueing delay as its
// duration.
func (s *KnowledgePostProcessService) finishRunningMultimodalStage(
	ctx context.Context,
	knowledgeID string,
	attempt int,
) {
	mm := s.tracker().LookupStage(ctx, knowledgeID, attempt, types.StageMultimodal)
	if mm == nil ||
		mm.Kind != types.SpanKindStage ||
		mm.Status != types.SpanStatusRunning {
		return
	}
	s.tracker().EndSpan(ctx, mm, nil)
}

// Handle implements asynq handler for TypeKnowledgePostProcess.
func (s *KnowledgePostProcessService) Handle(ctx context.Context, task *asynq.Task) error {
	var payload types.KnowledgePostProcessPayload
	if err := json.Unmarshal(task.Payload(), &payload); err != nil {
		return fmt.Errorf("unmarshal knowledge post process payload: %w", err)
	}

	logger.Infof(ctx, "[KnowledgePostProcess] Orchestrating post processing for knowledge: %s", payload.KnowledgeID)

	ctx = types.WithExecutionTenant(ctx, payload.TenantID)
	if payload.Language != "" {
		ctx = context.WithValue(ctx, types.LanguageContextKey, payload.Language)
	}

	// 1. Fetch Knowledge and KB
	knowledge, err := s.knowledgeRepo.GetKnowledgeByIDOnly(ctx, payload.KnowledgeID)
	if err != nil {
		return fmt.Errorf("get knowledge %s: %w", payload.KnowledgeID, err)
	}
	if knowledge == nil {
		logger.Warnf(ctx, "[KnowledgePostProcess] Knowledge %s not found, aborting.", payload.KnowledgeID)
		return nil
	}

	if err := validateProcessingKnowledge(knowledge,
		payload.TenantID,
		payload.KnowledgeBaseID,
		payload.KnowledgeID); err != nil {
		return err
	}
	kb, err := s.kbService.GetKnowledgeBaseByIDOnly(ctx, payload.KnowledgeBaseID)
	if err != nil || kb == nil {
		return fmt.Errorf("get knowledge base %s: %w", payload.KnowledgeBaseID, err)
	}

	if kb.ID != payload.KnowledgeBaseID || kb.TenantID != payload.TenantID {
		return fmt.Errorf("postprocess task KB owner changed: %w", asynq.SkipRetry)
	}
	ctx, err = access.WithKBTaskWrite(ctx, kb, payload.TenantID)
	if err != nil {
		return fmt.Errorf("invalid postprocess scope: %v: %w", err, asynq.SkipRetry)
	}
	// Resolve attempt: payload carries it from the upstream stage, but
	// fall back to the latest known attempt for compatibility with
	// in-flight tasks queued before this code shipped.
	attempt := payload.Attempt
	if attempt <= 0 {
		attempt = s.tracker().LatestAttempt(ctx, payload.KnowledgeID)
	} else if attemptSuperseded(ctx, s.tracker(), payload.KnowledgeID, attempt) {
		// A reparse started after this run. Entering finalizing here would
		// seed the counter for subtasks that all drop themselves as
		// superseded, and the new run's own post-process would then find
		// the row already finalizing and skip its fan-out.
		logger.Infof(ctx, "[KnowledgePostProcess] Attempt %d of %s superseded, skipping.",
			attempt, payload.KnowledgeID)
		return nil
	}

	// Close the multimodal stage span (parent enqueued it as "running"
	// 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 must remain so.
	// 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.
	s.finishRunningMultimodalStage(ctx, payload.KnowledgeID, attempt)

	postSpan := s.tracker().BeginStage(ctx, payload.KnowledgeID, attempt, types.StagePostProcess, nil)

	// Skip post-processing entirely when the knowledge has been cancelled
	// by the user or marked for deletion. We must NOT enqueue summary /
	// question / graph / wiki child tasks for an aborted knowledge. We
	// MUST also close postSpan before returning, otherwise it stays in
	// running state forever and the trace viewer shows an orange bar
	// long after the user cancelled (the AbortAttempt sweep ran before
	// we opened postSpan, so the sweep didn't catch this row).
	switch knowledge.ParseStatus {
	case types.ParseStatusCancelled, types.ParseStatusDeleting:
		logger.Infof(ctx,
			"[KnowledgePostProcess] Knowledge %s aborted (%s), skipping post-processing.",
			payload.KnowledgeID, knowledge.ParseStatus,
		)
		s.tracker().SkipSpan(ctx, postSpan,
			"knowledge "+knowledge.ParseStatus+" before postprocess started")
		return nil
	}

	processOverrides, _ := knowledge.ProcessOverrides()
	eff := ResolveProcessConfig(kb, processOverrides)

	// 2. Fetch all text-like chunks. ListChunksByKnowledgeID is text-only by
	//    design, which silently dropped the OCR / Caption chunks that
	//    multimodal parsing adds for scanned PDFs — leaving graph extraction
	//    with nothing but image-link placeholders. Ask the repository for the
	//    exact chunk types we enrich instead.
	chunks, err := s.chunkRepo.ListChunksByKnowledgeIDAndTypes(ctx, payload.TenantID, payload.KnowledgeID,
		[]types.ChunkType{types.ChunkTypeText, types.ChunkTypeImageOCR, types.ChunkTypeImageCaption})
	if err != nil {
		return fmt.Errorf("list chunks for knowledge %s: %w", payload.KnowledgeID, err)
	}

	for _, chunk := range chunks {
		if chunk == nil || chunk.TenantID != payload.TenantID || chunk.KnowledgeID != payload.KnowledgeID ||
			chunk.KnowledgeBaseID != payload.KnowledgeBaseID {
			return fmt.Errorf("postprocess chunk binding changed: %w", asynq.SkipRetry)
		}
	}

	// Gather all text-like chunks (including newly added OCR and Caption from multimodal tasks)
	var textChunks []*types.Chunk
	for _, c := range chunks {
		if c.ChunkType == types.ChunkTypeText || c.ChunkType == types.ChunkTypeImageOCR || c.ChunkType == types.ChunkTypeImageCaption {
			textChunks = append(textChunks, c)
		}
	}

	graphChunks := selectGraphChunks(textChunks)

	// 3. Compute the enrichment subtask count up front so we can flip to
	//    "finalizing" with the right counter BEFORE spawning any subtasks.
	//    Each subtask handler atomically decrements pending_subtasks_count
	//    on its terminal exit; the row promotes itself to "completed" when
	//    the counter hits zero (see knowledgeRepository.FinalizeSubtask).
	//
	//    Wiki ingest IS counted (as a single subtask): although it's a
	//    KB-scoped debounced batch, each upload enqueues exactly one
	//    per-knowledge op, and the batch worker calls FinalizeSubtask once
	//    when that op reaches a terminal state (mapped successfully or
	//    dead-lettered after exhausting retries). Counting it keeps the row
	//    in "finalizing" — i.e. shown as in-progress and still cancellable —
	//    until wiki generation actually finishes instead of flipping to
	//    completed while wiki runs minutes later. A wiki op that never
	//    drains is bounded by the housekeeping finalizing sweep.
	willSpawnSummary := eff.SummaryEnabled && len(textChunks) > 0
	willSpawnQuestion := len(textChunks) > 0 && kb.NeedsEmbeddingModel() &&
		eff.QuestionGenerationConfig.Enabled
	willSpawnWiki := kb.IndexingStrategy.WikiEnabled && len(textChunks) > 0
	willSpawnAutoTag := kb.Type == types.KnowledgeBaseTypeDocument &&
		kb.AutoTagConfig != nil && kb.AutoTagConfig.Enabled && len(textChunks) > 0
	enqueuedAutoTag := false

	// Question generation now fans out one subtask per plain text chunk
	// (mirroring the graph-extract per-chunk pattern) so each chunk's LLM
	// call retries / cancels / traces independently. We only target
	// ChunkTypeText here — OCR / Caption chunks were never fed to question
	// generation in the legacy whole-knowledge loop, so excluding them
	// keeps behavior identical. Link-only text chunks (scanned PDF pages)
	// are skipped: the LLM has nothing to ask about and tends to echo the
	// few-shot example. Sorted by StartAt so the per-chunk context
	// (prev / next) matches the legacy ordering.
	var questionChunks []*types.Chunk
	if willSpawnQuestion {
		for _, c := range textChunks {
			if c.ChunkType == types.ChunkTypeText && chunkHasExtractableText(c.Content) {
				questionChunks = append(questionChunks, c)
			}
		}
		sort.Slice(questionChunks, func(i, j int) bool {
			return questionChunks[i].StartAt < questionChunks[j].StartAt
		})
	}

	// Question generation is batched: one subtask per window of
	// questionGenChunkBatchSize text chunks (not one per chunk), so a
	// huge document doesn't spawn thousands of tiny tasks. The counter
	// must match exactly how many batch tasks we enqueue below.
	questionBatchCount := (len(questionChunks) + questionGenChunkBatchSize - 1) / questionGenChunkBatchSize

	graphChunkCount := 0
	if eff.GraphEnabled {
		graphChunkCount = len(graphChunks)
	}
	expectedSubtasks := 0
	if willSpawnSummary {
		expectedSubtasks++
	}
	expectedSubtasks += questionBatchCount
	if willSpawnWiki {
		expectedSubtasks++
	}
	expectedSubtasks += graphChunkCount

	// enteredFinalizing is set only when the processing-to-finalizing handoff
	// actually seeded the counter. For Wiki-enabled knowledge, that handoff
	// also persists the pending Wiki op in the same transaction.
	enteredFinalizing := false
	wikiSlotOwned := false

	switch {
	case knowledge.ParseStatus == types.ParseStatusFinalizing && kb.IndexingStrategy.WikiEnabled:
		// A previous delivery may have persisted the Wiki op but failed to
		// enqueue its KB-scoped trigger. Retry only the trigger: appending a
		// second pending op would duplicate durable work and its finalizer.
		if err := enqueueWikiIngestTrigger(ctx, s.taskEnqueuer, payload.TenantID, payload.KnowledgeBaseID); err != nil {
			s.tracker().FailSpanRetryable(ctx, postSpan, "WIKI_TRIGGER_ENQUEUE_FAILED", err.Error(), err)
			return fmt.Errorf("retry wiki ingest trigger: %w", err)
		}
		logger.Infof(ctx, "[KnowledgePostProcess] Re-enqueued wiki ingest trigger for %s", payload.KnowledgeID)
		retryOutput := types.JSONMap{"retried_wiki_trigger": true}
		s.tracker().EndSpan(ctx, postSpan, retryOutput)
		s.tracker().FinalizeAttempt(ctx, payload.KnowledgeID, attempt,
			types.SpanStatusDone, retryOutput, "", "")
		return nil
	case knowledge.ParseStatus != types.ParseStatusProcessing:
		// The row was already in some other state (deleting / cancelled /
		// failed / completed) when we arrived. Don't touch parse_status
		// and don't spawn enrichment — the upstream that put the row in
		// that state has already decided this attempt is over.
		logger.Infof(ctx, "[KnowledgePostProcess] Knowledge %s is in %s, skipping enrichment fan-out.",
			payload.KnowledgeID, knowledge.ParseStatus)
		s.tracker().EndSpan(ctx, postSpan, types.JSONMap{
			"skipped":         "non_processing_status",
			"observed_status": knowledge.ParseStatus,
		})
		s.tracker().FinalizeAttempt(ctx, payload.KnowledgeID, attempt,
			types.SpanStatusDone, types.JSONMap{
				"skipped":         "non_processing_status",
				"observed_status": knowledge.ParseStatus,
			}, "", "")
		return nil
	case expectedSubtasks == 0:
		completed, err := s.knowledgeRepo.CompleteProcessingWithoutSubtasks(ctx, payload.KnowledgeID)
		if err != nil {
			s.tracker().FailSpanRetryable(ctx, postSpan, "COMPLETION_FAILED", err.Error(), err)
			return fmt.Errorf("complete knowledge without enrichment: %w", err)
		}
		if !completed {
			output := types.JSONMap{"skipped": "knowledge_no_longer_processing"}
			s.tracker().EndSpan(ctx, postSpan, output)
			s.tracker().FinalizeAttempt(ctx, payload.KnowledgeID, attempt, types.SpanStatusDone, output, "", "")
			return nil
		}
	default:
		// Flip processing to finalizing before fan-out so a parallel
		// cancel/delete cannot race us into completed.
		var promoted bool
		var err error
		if willSpawnWiki {
			seeder, ok := s.pendingRepo.(interfaces.TaskPendingOpsFinalizingSeeder)
			if !ok {
				return errors.New("wiki post-process requires atomic finalizing handoff")
			}
			pendingOp, buildErr := newWikiIngestPendingOp(
				withAttempt(ctx, attempt), payload.TenantID, payload.KnowledgeBaseID, payload.KnowledgeID,
			)
			if buildErr != nil {
				return buildErr
			}
			promoted, err = seeder.SeedKnowledgeFinalizingWithPendingOp(
				ctx, payload.KnowledgeID, expectedSubtasks, pendingOp,
			)
			wikiSlotOwned = promoted
		} else {
			promoted, err = s.knowledgeRepo.SetFinalizing(ctx, payload.KnowledgeID, expectedSubtasks)
		}
		if err != nil {
			// Acking here would read as "no longer processing" below and
			// strand the row; retry so the handoff (or dead-letter) happens.
			logger.Warnf(ctx, "[KnowledgePostProcess] SetFinalizing failed for %s: %v",
				payload.KnowledgeID, err)
			s.tracker().FailSpanRetryable(ctx, postSpan, "FINALIZING_HANDOFF_FAILED", err.Error(), err)
			return fmt.Errorf("enter finalizing: %w", err)
		}
		if promoted {
			enteredFinalizing = true
			// Reflect summary status separately so the UI shows the
			// summary as queued for users who already had it visible.
			summaryStatus := types.SummaryStatusNone
			if willSpawnSummary {
				summaryStatus = types.SummaryStatusPending
			}
			if err := s.knowledgeRepo.UpdateKnowledgeColumn(ctx,
				payload.KnowledgeID, "summary_status", summaryStatus); err != nil {
				logger.Warnf(ctx, "[KnowledgePostProcess] Failed to update summary_status for %s: %v",
					payload.KnowledgeID, err)
			}
			logger.Infof(ctx,
				"[KnowledgePostProcess] Knowledge %s entered finalizing (pending_subtasks=%d).",
				payload.KnowledgeID, expectedSubtasks)
		} else {
			// Row was no longer 'processing' (cancel / delete won the race).
			// Skip enrichment entirely so we don't waste LLM quota on a row
			// the user already abandoned.
			logger.Infof(ctx,
				"[KnowledgePostProcess] Knowledge %s no longer in processing, skipping enrichment fan-out.",
				payload.KnowledgeID)
			s.tracker().EndSpan(ctx, postSpan, types.JSONMap{
				"skipped": "knowledge_no_longer_processing",
			})
			s.tracker().FinalizeAttempt(ctx, payload.KnowledgeID, attempt,
				types.SpanStatusDone, types.JSONMap{
					"skipped": "knowledge_no_longer_processing",
				}, "", "")
			return nil
		}
	}

	// Queue best-effort automatic tagging only after the processing row has
	// successfully handed off to finalizing or completed. This avoids model calls from a
	// duplicate post-process delivery that observes an already terminal row.
	if willSpawnAutoTag {
		enqueuedAutoTag = s.enqueueAutoTagTask(ctx, payload, attempt)
	}

	// 4. Spawn Summary and Question Tasks
	enqueuedSummary := false
	enqueuedQuestionCount := 0
	if willSpawnSummary {
		enqueuedSummary = s.enqueueSummaryGenerationTask(ctx, payload, attempt)
		if !enqueuedSummary {
			_ = s.knowledgeRepo.UpdateKnowledgeColumn(
				ctx, payload.KnowledgeID, "summary_status", types.SummaryStatusFailed,
			)
		}
	}
	if willSpawnQuestion {
		// Create the postprocess.question grouping span up front so the
		// per-batch subspans (enqueued just below, run later in their own
		// workers) have a parent to nest under. It's begun and ended right
		// here as a structural container — the batches extend past it,
		// which the timeline renders with the wrapping outline bar.
		if grp := s.tracker().BeginSubSpan(ctx, postSpan, postprocessQuestionGroupSpanName,
			types.SpanKindSubSpan, types.JSONMap{
				"batch_count": questionBatchCount,
				"chunk_count": len(questionChunks),
				"batch_size":  questionGenChunkBatchSize,
			}); grp != nil {
			s.tracker().EndSpan(ctx, grp, types.JSONMap{
				"batch_count": questionBatchCount,
				"chunk_count": len(questionChunks),
			})
		}
		enqueuedQuestionCount = s.enqueueQuestionGenerationTasks(
			ctx, payload, eff.QuestionGenerationConfig, attempt, questionChunks,
		)
	}

	// 5. Spawn Graph RAG Tasks — only when graph indexing is enabled in IndexingStrategy
	enqueuedGraphCount := 0
	if graphChunkCount > 0 {
		logger.Infof(ctx,
			"[KnowledgePostProcess] Spawning Graph RAG extract tasks for %d text/OCR chunks (of %d text-like)",
			len(graphChunks), len(textChunks))
		for i, chunk := range graphChunks {
			ok, err := NewChunkExtractTask(ctx, s.taskEnqueuer, payload.TenantID, chunk.ID, kb.SummaryModelID,
				payload.KnowledgeID, attempt, i)
			if err != nil {
				logger.Errorf(ctx, "[KnowledgePostProcess] Failed to create chunk extract task for %s: %v", chunk.ID, err)
			}
			if ok {
				enqueuedGraphCount++
			}
		}
	}

	// 6. Schedule the Wiki trigger. The durable per-knowledge op already owns
	//    its finalizing slot because it was committed atomically with the state
	//    transition above. A trigger failure is returned so the post-process
	//    task retries only the trigger without double-accounting.
	var wikiEnqueueErr error
	if willSpawnWiki {
		wikiEnqueueErr = enqueueWikiIngestTrigger(
			ctx, s.taskEnqueuer, payload.TenantID, payload.KnowledgeBaseID,
		)
		if wikiEnqueueErr != nil {
			logger.Warnf(ctx, "[KnowledgePostProcess] Failed to enqueue wiki ingest for %s: %v",
				payload.KnowledgeID, wikiEnqueueErr)
		} else if wikiSlotOwned {
			logger.Infof(ctx, "[KnowledgePostProcess] Enqueued wiki ingest task for %s", payload.KnowledgeID)
		}
	}

	// Reconcile the seeded counter against what was actually enqueued.
	// summary/question/graph each own a counted slot that ONLY their own
	// task drains; a slot whose task was never enqueued (graph with NEO4J
	// off, a transient enqueue/marshal failure, a nil enqueuer) has no owner
	// and would otherwise strand the row in "finalizing". Release exactly the
	// shortfall — each release is a clamped decrement that promotes the row to
	// "completed" if it brings the counter to zero. Safe against fast workers:
	// shortfall slots have no draining
	// task, so total drains == seeded count regardless of ordering.
	//
	// Detached ctx: the same reasoning that motivates finalizeSubtaskDetached
	// for terminal worker drains applies here. If the postprocess handler's
	// ctx is cancelled (graceful shutdown, preempted worker) between SetFinalizing
	// and this point, the seeded slots have NO other path to drain — every
	// owning task either failed to enqueue or was never created. Riding a
	// cancelled ctx would silently abort the releases and strand the row in
	// "finalizing". The bound is per-call (matches the helper) so a wedged
	// connection can't pin the goroutine for the whole serial loop.
	if enteredFinalizing {
		plannedOwned := questionBatchCount + graphChunkCount
		if willSpawnSummary {
			plannedOwned++
		}
		if willSpawnWiki {
			plannedOwned++
		}
		actualOwned := enqueuedQuestionCount + enqueuedGraphCount
		if enqueuedSummary {
			actualOwned++
		}
		if wikiSlotOwned {
			actualOwned++
		}
		if shortfall := plannedOwned - actualOwned; shortfall > 0 {
			logger.Warnf(ctx,
				"[KnowledgePostProcess] Releasing %d un-enqueued subtask slot(s) for %s (planned=%d actual=%d)",
				shortfall, payload.KnowledgeID, plannedOwned, actualOwned)
			// Keep going past a failed release: stopping at the first error
			// left every remaining slot without an owner.
			for i := 0; i < shortfall; i++ {
				if err := releaseSubtaskSlot(ctx, s.knowledgeRepo, payload.KnowledgeID); err != nil {
					logger.Errorf(ctx, "[KnowledgePostProcess] Failed to release subtask slot for %s: %v",
						payload.KnowledgeID, err)
				}
			}
		}
	}

	postOutput := types.JSONMap{
		"chunks_total":            len(textChunks),
		"enqueued_summary":        enqueuedSummary,
		"enqueued_question":       enqueuedQuestionCount > 0,
		"enqueued_question_count": enqueuedQuestionCount,
		"enqueued_wiki":           wikiSlotOwned && wikiEnqueueErr == nil,
		"wiki_slot_owned":         wikiSlotOwned,
		"enqueued_graph":          enqueuedGraphCount > 0,
		"enqueued_graph_count":    enqueuedGraphCount,
		"enqueued_auto_tag":       enqueuedAutoTag,
	}
	s.tracker().EndSpan(ctx, postSpan, postOutput)
	if wikiSlotOwned && wikiEnqueueErr != nil {
		return fmt.Errorf("enqueue wiki ingest trigger: %w", wikiEnqueueErr)
	}
	// 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, "", "")
	// The document now counts (title, type, folder, tags) in the knowledge-base
	// description aggregation even before its summary lands; the summary task
	// requests another refresh once the profile exists. No-op unless the KB
	// opted in, and debounced so a batch upload costs one aggregation.
	_ = requestKnowledgeBaseProfileRefresh(ctx, s.taskEnqueuer, kb, false)
	return nil
}

// enqueueAutoTagTask schedules best-effort classification against the KB's
// existing tags. It intentionally owns no pending-subtask slot: a model or
// configuration failure must never keep document parsing in finalizing.
func (s *KnowledgePostProcessService) enqueueAutoTagTask(
	ctx context.Context,
	payload types.KnowledgePostProcessPayload,
	attempt int,
) bool {
	if s.taskEnqueuer == nil {
		return false
	}
	taskPayload := types.KnowledgeAutoTagPayload{
		TenantID:        payload.TenantID,
		KnowledgeBaseID: payload.KnowledgeBaseID,
		KnowledgeID:     payload.KnowledgeID,
		Language:        payload.Language,
		Attempt:         attempt,
	}
	langfuse.InjectTracing(ctx, &taskPayload)
	payloadBytes, err := json.Marshal(taskPayload)
	if err != nil {
		logger.Warnf(ctx, "[KnowledgePostProcess] Failed to marshal auto tag payload: %v", err)
		return false
	}
	task := asynq.NewTask(types.TypeKnowledgeAutoTag, payloadBytes,
		asynq.Queue(types.QueueSummary), asynq.MaxRetry(2), asynq.Timeout(2*time.Minute))
	if _, err := s.taskEnqueuer.Enqueue(task); err != nil {
		logger.Warnf(ctx, "[KnowledgePostProcess] Failed to enqueue automatic tagging for %s: %v", payload.KnowledgeID, err)
		return false
	}
	logger.Infof(ctx, "[KnowledgePostProcess] Enqueued automatic tagging for %s", payload.KnowledgeID)
	return true
}

// enqueueSummaryGenerationTask enqueues the summary task. Returns true only
// when a task was actually placed on the queue, so the caller can release the
// seeded pending-subtask slot when enqueue is skipped or fails.
func (s *KnowledgePostProcessService) enqueueSummaryGenerationTask(ctx context.Context, payload types.KnowledgePostProcessPayload, attempt int) bool {
	if s.taskEnqueuer == nil {
		return false
	}

	taskPayload := types.SummaryGenerationPayload{
		TenantID:        payload.TenantID,
		KnowledgeBaseID: payload.KnowledgeBaseID,
		KnowledgeID:     payload.KnowledgeID,
		Language:        payload.Language,
		Attempt:         attempt,
	}
	langfuse.InjectTracing(ctx, &taskPayload)
	payloadBytes, err := json.Marshal(taskPayload)
	if err != nil {
		logger.Warnf(ctx, "[KnowledgePostProcess] Failed to marshal summary generation payload: %v", err)
		return false
	}

	task := asynq.NewTask(types.TypeSummaryGeneration, payloadBytes,
		asynq.Queue(types.QueueSummary), asynq.MaxRetry(3), asynq.Timeout(30*time.Minute))
	if _, err := s.taskEnqueuer.Enqueue(task); err != nil {
		logger.Warnf(ctx, "[KnowledgePostProcess] Failed to enqueue summary generation for %s: %v", payload.KnowledgeID, err)
		return false
	}
	logger.Infof(ctx, "[KnowledgePostProcess] Enqueued summary generation task for %s", payload.KnowledgeID)
	return true
}

// questionGenChunkBatchSize is the number of text chunks handled by a single
// question-generation task. Batching keeps the task count bounded for very
// large documents (a 5k-chunk doc becomes ~250 tasks instead of 5k) while
// preserving per-batch retry / cancellation granularity and letting each task
// do one embedding BatchIndex over the whole batch.
const questionGenChunkBatchSize = 20

// postprocessQuestionGroupSpanName is the grouping span the per-batch
// question subspans (postprocess.question.batch[i]) nest under, so the trace
// viewer shows one "postprocess.question" node instead of dozens of siblings
// directly beneath the postprocess stage.
const postprocessQuestionGroupSpanName = "postprocess.question"

// enqueueQuestionGenerationTasks fans out one TypeQuestionGeneration task per
// batch of questionGenChunkBatchSize text chunks. Each task carries only chunk
// ids (+ the adjacent boundary ids for context) — never the chunk content — so
// the payload stays small and the worker reads fresh content at run time,
// matching the ExtractChunkPayload precedent.
//
// Returns the number of batch tasks successfully enqueued. A failed
// marshal/enqueue is logged and skipped; the caller's reconciliation
// step (the shortfall-release loop in Handle) compares this count
// against questionBatchCount and releases any unowned slots so a
// half-fanned-out batch can't strand the row in "finalizing".
func (s *KnowledgePostProcessService) enqueueQuestionGenerationTasks(
	ctx context.Context,
	payload types.KnowledgePostProcessPayload,
	qg types.QuestionGenerationConfig,
	attempt int,
	questionChunks []*types.Chunk,
) int {
	if s.taskEnqueuer == nil || len(questionChunks) == 0 {
		return 0
	}
	if !qg.Enabled {
		return 0
	}

	questionCount := qg.QuestionCount
	if questionCount <= 0 {
		questionCount = 3
	}
	if questionCount > 10 {
		questionCount = 10
	}

	total := len(questionChunks)
	enqueued := 0
	batchIndex := 0
	for start := 0; start < total; start += questionGenChunkBatchSize {
		end := start + questionGenChunkBatchSize
		if end > total {
			end = total
		}
		batch := questionChunks[start:end]
		chunkIDs := make([]string, len(batch))
		for i, c := range batch {
			chunkIDs[i] = c.ID
		}

		taskPayload := types.QuestionGenerationPayload{
			TenantID:        payload.TenantID,
			KnowledgeBaseID: payload.KnowledgeBaseID,
			KnowledgeID:     payload.KnowledgeID,
			QuestionCount:   questionCount,
			Language:        payload.Language,
			Attempt:         attempt,
			ChunkIDs:        chunkIDs,
			BatchIndex:      batchIndex,
		}
		// Boundary context: the text chunk just before / after this window.
		if start > 0 {
			taskPayload.PrevChunkID = questionChunks[start-1].ID
		}
		if end < total {
			taskPayload.NextChunkID = questionChunks[end].ID
		}
		batchIndex++

		langfuse.InjectTracing(ctx, &taskPayload)
		payloadBytes, err := json.Marshal(taskPayload)
		if err != nil {
			logger.Warnf(ctx, "[KnowledgePostProcess] Failed to marshal question generation payload for batch %d: %v", batchIndex-1, err)
			continue
		}

		task := asynq.NewTask(types.TypeQuestionGeneration, payloadBytes,
			asynq.Queue(types.QueueQuestion), asynq.MaxRetry(3), asynq.Timeout(30*time.Minute))
		if _, err := s.taskEnqueuer.Enqueue(task); err != nil {
			logger.Warnf(ctx, "[KnowledgePostProcess] Failed to enqueue question generation batch %d for %s: %v", batchIndex-1, payload.KnowledgeID, err)
			continue
		}
		enqueued++
	}
	logger.Infof(ctx, "[KnowledgePostProcess] Enqueued %d question generation batch tasks (%d chunks, batch_size=%d) for %s (count=%d)",
		enqueued, total, questionGenChunkBatchSize, payload.KnowledgeID, questionCount)
	return enqueued
}

// selectGraphChunks picks the chunks that should be sent to graph extraction.
// Captions are dropped (they describe the page visually and duplicate OCR).
// Text chunks that are only image placeholders are skipped so the extractor
// cannot echo the few-shot example. OCR is kept when the parent text chunk
// has no extractable prose (scanned PDF pages); OCR of figures next to real
// text is skipped to avoid doubling LLM cost on illustrated documents.
func selectGraphChunks(chunks []*types.Chunk) []*types.Chunk {
	textByID := make(map[string]*types.Chunk, len(chunks))
	for _, c := range chunks {
		if c != nil && c.ChunkType == types.ChunkTypeText {
			textByID[c.ID] = c
		}
	}
	var out []*types.Chunk
	for _, c := range chunks {
		if c == nil {
			continue
		}
		switch c.ChunkType {
		case types.ChunkTypeImageCaption:
			continue
		case types.ChunkTypeImageOCR:
			if !chunkHasExtractableText(c.Content) {
				continue
			}
			if parent := textByID[c.ParentChunkID]; parent != nil && chunkHasExtractableText(parent.Content) {
				continue
			}
			out = append(out, c)
		case types.ChunkTypeText:
			if chunkHasExtractableText(c.Content) {
				out = append(out, c)
			}
		}
	}
	return out
}

// chunkHasExtractableText reports whether a chunk still contains prose once
// markdown images and <image> wrappers are removed, i.e. whether graph
// extraction (or question generation) can find entities in it.
func chunkHasExtractableText(content string) bool {
	return extractRealText(docparser.StripMarkdownImages(content)) != ""
}