Skip to content

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

30.5 KB
Raw Blame
package service

import (
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"strings"

	"github.com/Tencent/WeKnora/internal/application/service/retriever"
	apperrors "github.com/Tencent/WeKnora/internal/errors"
	"github.com/Tencent/WeKnora/internal/logger"
	"github.com/Tencent/WeKnora/internal/types"
	"github.com/Tencent/WeKnora/internal/types/interfaces"
	"github.com/hibiken/asynq"
	"golang.org/x/sync/errgroup"
)

// collectImageURLs extracts unique provider:// image URLs from image_info JSON strings.
func collectImageURLs(ctx context.Context, imageInfos []string) []string {
	seen := make(map[string]struct{})
	var urls []string
	for _, info := range imageInfos {
		if info == "" {
			continue
		}
		var images []*types.ImageInfo
		if err := json.Unmarshal([]byte(info), &images); err != nil {
			logger.Warnf(ctx, "Failed to parse image_info JSON: %v", err)
			continue
		}
		for _, img := range images {
			if img.URL != "" {
				if _, exists := seen[img.URL]; !exists {
					seen[img.URL] = struct{}{}
					urls = append(urls, img.URL)
				}
			}
		}
	}
	return urls
}

// mergeKnowledgeReleaseURLs unions ImageInfo URLs with derived catalog
// bindings for the knowledge entries being removed. Parsed documents often
// embed resource:// handles only in markdown; multimodal ImageInfo is empty
// until (or unless) those tasks finish, so attachment/extracted-image rows
// are the complete claim set. Source files stay out of this list: reparse
// and manual cleanup must keep the original document, and knowledge delete
// already removes FilePath separately.
//
// Only call this from a path that is removing the knowledge for good
// (executeKnowledgeDelete, ProcessKBDelete). Cleanup that runs before a
// re-index must keep collecting ImageInfo alone: the knowledge survives it,
// and DeleteFile marks the object deleted, so the re-claim that follows could
// not resolve the handles it dropped.
func mergeKnowledgeReleaseURLs(
	ctx context.Context,
	catalog interfaces.ResourceCatalog,
	knowledgeIDs []string,
	imageURLs []string,
) []string {
	seen := make(map[string]struct{}, len(imageURLs))
	out := make([]string, 0, len(imageURLs))
	add := func(url string) {
		url = strings.TrimSpace(url)
		if url == "" {
			return
		}
		if _, exists := seen[url]; exists {
			return
		}
		seen[url] = struct{}{}
		out = append(out, url)
	}
	for _, url := range imageURLs {
		add(url)
	}
	if catalog == nil || len(knowledgeIDs) == 0 {
		return out
	}
	refs, err := catalog.ListReferencesByOwner(ctx, types.ResourceOwnerKnowledge, knowledgeIDs...)
	if err != nil {
		logger.Warnf(ctx, "Failed to list knowledge resource bindings for release: %v", err)
		return out
	}
	for _, ref := range refs {
		add(ref)
	}
	return out
}

// knowledgeResourceOwners builds the releaser for a set of knowledge entries
// being deleted. A nil catalog (no resource registry) yields nil, which
// deleteExtractedImages treats as "delete unconditionally", i.e. the behaviour
// from before bindings were tracked.
func knowledgeResourceOwners(catalog interfaces.ResourceCatalog, knowledgeIDs ...string) *resourceReleaser {
	if catalog == nil || len(knowledgeIDs) == 0 {
		return nil
	}
	return &resourceReleaser{catalog: catalog, ownerType: types.ResourceOwnerKnowledge, ownerIDs: knowledgeIDs}
}

// resourceReleaser decides whether a stored file's bytes may be deleted along
// with the domain object that referenced them.
//
// A file can be claimed by more than one owner — an answer saved into the
// knowledge base shares the very blob the chat message still shows, and two
// knowledge entries can share an image. Deleting one owner must drop only that
// owner's claim; the bytes go away when the last claim does.
type resourceReleaser struct {
	catalog   interfaces.ResourceCatalog
	ownerType string
	ownerIDs  []string
}

// deletable drops this releaser's claims on ref and reports whether the bytes
// are now unreferenced.
//
// An unreadable binding count keeps the file. The cost of keeping it is an
// orphaned blob that a later delete can still reclaim; the cost of guessing
// wrong the other way is an image vanishing from a conversation or document
// that nobody deleted.
func (r *resourceReleaser) deletable(ctx context.Context, ref string) bool {
	if r == nil || r.catalog == nil {
		return true
	}
	remaining := int64(-1)
	for _, ownerID := range r.ownerIDs {
		count, err := r.catalog.Release(ctx, ref, r.ownerType, ownerID)
		if err != nil {
			logger.Warnf(ctx, "Failed to release resource %s from %s %s: %v",
				ref, r.ownerType, ownerID, err)
			return false
		}
		if count >= 0 {
			remaining = count
		}
	}
	// -1 means the reference is not a catalog handle: no claims to account
	// for, so fall back to deleting it as before.
	return remaining <= 0
}

// deleteExtractedImages deletes all extracted image files from storage.
// Standalone function — callable from both knowledgeService and knowledgeBaseService.
// Errors are logged but do not fail the overall deletion.
//
// releaser may be nil, in which case every file is deleted unconditionally.
func deleteExtractedImages(
	ctx context.Context,
	fileSvc interfaces.FileService,
	releaser *resourceReleaser,
	imageURLs []string,
) {
	if len(imageURLs) == 0 {
		return
	}
	logger.Infof(ctx, "Deleting %d extracted images", len(imageURLs))
	for _, url := range imageURLs {
		if !releaser.deletable(ctx, url) {
			logger.Infof(ctx, "Keeping extracted image %s: another owner may still reference it", url)
			continue
		}
		if err := fileSvc.DeleteFile(ctx, url); err != nil {
			logger.Errorf(ctx, "Failed to delete extracted image %s: %v", url, err)
		}
	}
}

// DeleteKnowledge deletes a knowledge entry and all related resources
func (s *knowledgeService) DeleteKnowledge(ctx context.Context, id string) error {
	plan, err := s.planKnowledgeDelete(ctx, []string{id})
	if err != nil {
		return err
	}
	return s.executeKnowledgeDelete(plan, true)
}

// cleanupWikiOnKnowledgeDelete handles wiki pages when a source document is deleted.
//
// There are three sources of truth we must keep consistent:
//   - The knowledge row (being soft-deleted right now by the caller)
//   - Wiki pages whose source_refs include this knowledge
//   - Pending/in-flight wiki_ingest tasks that may create *new* pages pointing at it
//
// The function is deliberately best-effort and idempotent:
//   - It writes a tombstone + scrubs pending ingest ops so new pages cannot be
//     born with a stale source_ref (guards (a) queued ingest and (b) ingest
//     tasks mid-LLM call — both consult the tombstone before writing).
//   - It persists a retract task before reconciling existing pages (delete
//     if only source, or strip the source if shared), preserving retry evidence.
//     Crucially we DO NOT gate
//     enqueue on "pages currently exist": in the ingest/delete race the
//     knowledge may have pages that exist only after this function returns
//     (the ingest task fires later and, absent the tombstone, would have
//     created them). The retract handler re-queries ListPagesBySourceRef at
//     run time, so even with an empty PageSlugs it will do the right thing —
//     and at worst it's a cheap no-op.
func (s *knowledgeService) cleanupWikiOnKnowledgeDelete(ctx context.Context, knowledge *types.Knowledge) {
	if err := s.cleanupWikiReferences(ctx, knowledge, s.wikiChunkRefsForKnowledge(ctx, knowledge)); err != nil {
		logger.Warnf(ctx, "wiki cleanup failed: %v", err)
	}
}

func (s *knowledgeService) cleanupWikiReferences(
	ctx context.Context,
	knowledge *types.Knowledge,
	sourceChunkRefs map[string]bool,
) error {
	if knowledge == nil {
		return nil
	}
	kbID := knowledge.KnowledgeBaseID
	knowledgeID := knowledge.ID
	if kbID == "" || knowledgeID == "" {
		return nil
	}

	var cleanupErr error

	// (1) Tombstone + scrub pending ingest — must happen first so any
	// wiki_ingest task that wakes up between here and the retract enqueue
	// below sees "knowledge gone" and bails out.
	if s.redisClient != nil {
		key := WikiDeletedTombstoneKey(kbID, knowledgeID)
		if err := s.redisClient.Set(ctx, key, "1", wikiDeletedTTL).Err(); err != nil {
			cleanupErr = errors.Join(cleanupErr, err)
		}
	}
	if s.taskPendingRepo != nil {
		err := s.taskPendingRepo.DeleteByDedupKey(ctx, wikiTaskType, wikiTaskScope, kbID, knowledgeID, WikiOpIngest)
		if err != nil {
			cleanupErr = errors.Join(cleanupErr, err)
		}
	}

	// Pull title/summary from the knowledge itself — do NOT read them from
	// existing wiki pages. In the race window wiki pages may not exist yet,
	// and even when they do their "summary" is the LLM-extracted one which
	// we're about to invalidate anyway. The knowledge row still has the
	// original Title/FileName/Description, which is what the retract prompt
	// actually wants.
	docTitle := knowledge.Title
	if docTitle == "" {
		docTitle = knowledge.FileName
	}
	if docTitle == "" {
		docTitle = knowledgeID
	}
	docSummary := knowledge.Description

	// (2) Immediate reconciliation for pages already present. If ingest
	// hasn't run yet this simply finds nothing; that's fine — see (3).
	pages, err := s.wikiRepo.ListBySourceRef(ctx, kbID, knowledgeID)
	if err != nil {
		logger.Warnf(ctx, "wiki cleanup: failed to list pages by source ref %s: %v", knowledgeID, err)
		return errors.Join(cleanupErr, err)
	}

	// Prefer the on-disk summary if the summary page already exists (it's
	// richer than the raw user-provided description). Leave docSummary
	// untouched otherwise so we still pass something meaningful downstream.
	for _, page := range pages {
		if page.PageType == types.WikiPageTypeSummary && page.Summary != "" {
			docSummary = page.Summary
			break
		}
	}

	// Persist the retract page set before removing source refs. Otherwise an
	// enqueue failure would leave the retry unable to rediscover those pages.
	var pageSlugs, folderIDs []string
	for _, page := range pages {
		if page.PageType != types.WikiPageTypeIndex {
			pageSlugs = append(pageSlugs, page.Slug)
			folderIDs = append(folderIDs, page.FolderID)
		}
	}
	lang := types.LanguageFromContextOrDefault(ctx)
	tenantID, _ := types.TenantIDFromContext(ctx)
	if err := enqueueWikiRetract(ctx, s.task, s.taskPendingRepo, WikiRetractPayload{
		TenantID: tenantID, KnowledgeBaseID: kbID, KnowledgeID: knowledgeID,
		DocTitle: docTitle, DocSummary: docSummary, Language: lang,
		PageSlugs: pageSlugs, FolderIDs: uniqueWikiFolderIDs(folderIDs),
	}); err != nil {
		return errors.Join(cleanupErr, err)
	}

	var deletedSlugs []string
	for _, page := range pages {
		if page.PageType == types.WikiPageTypeIndex {
			continue
		}

		remaining := removeSourceRef(page.SourceRefs, knowledgeID)

		if len(remaining) == 0 {
			if err := s.wikiService.DeletePage(ctx, kbID, page.Slug); err != nil {
				logger.Warnf(ctx, "wiki cleanup: failed to delete page %s: %v", page.Slug, err)
				cleanupErr = errors.Join(cleanupErr, err)
			} else {
				deletedSlugs = append(deletedSlugs, page.Slug)
			}
		} else {
			page.SourceRefs = remaining
			page.ChunkRefs = removeChunkRefs(page.ChunkRefs, sourceChunkRefs)
			if err := s.wikiService.UpdatePageMeta(ctx, page); err != nil {
				logger.Warnf(ctx, "wiki cleanup: failed to update source refs for page %s: %v", page.Slug, err)
				cleanupErr = errors.Join(cleanupErr, err)
			}
		}
	}

	if len(deletedSlugs) > 0 {
		logger.Infof(ctx, "wiki cleanup: deleted %d pages after knowledge %s deletion: %v",
			len(deletedSlugs), knowledgeID, deletedSlugs)
	}

	return cleanupErr
}

func (s *knowledgeService) wikiChunkRefsForKnowledge(ctx context.Context, knowledge *types.Knowledge) map[string]bool {
	if knowledge == nil || s.chunkRepo == nil {
		return nil
	}
	chunks, err := s.chunkRepo.ListAllChunksByKnowledgeID(ctx, knowledge.TenantID, knowledge.ID)
	if err != nil {
		logger.Warnf(ctx, "wiki cleanup: failed to list chunks for knowledge %s: %v", knowledge.ID, err)
		return nil
	}
	refs := make(map[string]bool, len(chunks))
	for _, chunk := range chunks {
		if chunk == nil || chunk.ID == "" {
			continue
		}
		refs[chunk.ID] = true
	}
	return refs
}

// scrubWikiPendingIngest removes queued WikiOpIngest entries for a knowledge
// from task_pending_ops. Used by both the delete path (we're about to
// soft-delete the doc, no point ingesting it) and the reparse path (the
// old chunks are about to vanish, so any pending ingest would either race
// with the cleanup or no-op on an empty chunk set — and the post-process
// task will enqueue a fresh ingest once new chunks land anyway).
//
// Retract entries stay put — delete still needs them to unlink referencing
// pages, and reparse never enqueues retracts for the doc being reparsed.
// We pass op=WikiOpIngest so DeleteByDedupKey filters to the ingest rows
// only.
func (s *knowledgeService) scrubWikiPendingIngest(ctx context.Context, kbID, knowledgeID, reason string) {
	if s.taskPendingRepo == nil || kbID == "" || knowledgeID == "" {
		return
	}
	err := s.taskPendingRepo.DeleteByDedupKey(ctx, wikiTaskType, wikiTaskScope, kbID, knowledgeID, WikiOpIngest)
	if err != nil {
		logger.Warnf(ctx, "wiki %s: failed to scrub pending ingest ops for knowledge %s: %v", reason, knowledgeID, err)
		return
	}
	logger.Infof(ctx, "wiki %s: scrubbed pending ingest ops for knowledge %s", reason, knowledgeID)
}

// prepareWikiForReparse is the reparse counterpart to
// cleanupWikiOnKnowledgeDelete. It aligns reparse with the same "pending
// queue hygiene" the delete path already enforces, without taking any
// destructive action against existing pages.
//
// Why no retract / tombstone here: reparse is not a "K is gone" event, it's
// a "K's contribution is about to be swapped for a new version" event. The
// actual swap happens asynchronously inside mapOneDocument (see its
// oldPageSlugs handling) — that's where we have both the old page set and
// the freshly extracted candidate slugs, which is exactly the information
// the WikiPageModifyUserPrompt needs to do a correct replace-not-append.
//
// So the only thing worth doing synchronously at reparse time is keeping
// the Redis pending list clean so the re-ingest enqueued by
// KnowledgePostProcess doesn't race with a stale ingest op that would
// fire mid-flight against zero chunks.
func (s *knowledgeService) prepareWikiForReparse(ctx context.Context, knowledge *types.Knowledge) {
	if knowledge == nil {
		return
	}
	kbID := knowledge.KnowledgeBaseID
	knowledgeID := knowledge.ID
	if kbID == "" || knowledgeID == "" {
		return
	}
	s.scrubWikiPendingIngest(ctx, kbID, knowledgeID, "reparse")
}

// removeSourceRef removes entries from source_refs that match a knowledge ID.
// Handles both old format ("knowledgeID") and new format ("knowledgeID|title").
func removeSourceRef(refs types.StringArray, knowledgeID string) types.StringArray {
	var result types.StringArray
	for _, ref := range refs {
		if types.WikiSourceRefMatchesKnowledge(ref, knowledgeID) {
			continue
		}
		result = append(result, ref)
	}
	return result
}

func removeChunkRefs(refs types.StringArray, removed map[string]bool) types.StringArray {
	if len(refs) == 0 || len(removed) == 0 {
		return refs
	}
	result := make(types.StringArray, 0, len(refs))
	for _, ref := range refs {
		if removed[ref] {
			continue
		}
		result = append(result, ref)
	}
	return result
}

type knowledgeVectorDeleteGroup struct {
	VectorStoreID    string
	EmbeddingModelID string
	Type             string
	KnowledgeIDs     []string
}

func buildKnowledgeVectorDeleteGroups(
	knowledges []*types.Knowledge,
	knowledgeBases map[string]*types.KnowledgeBase,
) []knowledgeVectorDeleteGroup {
	type groupKey struct {
		VectorStoreID    string
		EmbeddingModelID string
		Type             string
	}

	grouped := make(map[groupKey][]string)
	for _, knowledge := range knowledges {
		if knowledge == nil {
			continue
		}
		var vectorStoreID string
		if kb := knowledgeBases[knowledge.KnowledgeBaseID]; kb != nil && kb.VectorStoreID != nil {
			vectorStoreID = strings.TrimSpace(*kb.VectorStoreID)
		}
		key := groupKey{
			VectorStoreID:    vectorStoreID,
			EmbeddingModelID: knowledge.EmbeddingModelID,
			Type:             knowledge.Type,
		}
		grouped[key] = append(grouped[key], knowledge.ID)
	}

	groups := make([]knowledgeVectorDeleteGroup, 0, len(grouped))
	for key, knowledgeIDs := range grouped {
		groups = append(groups, knowledgeVectorDeleteGroup{
			VectorStoreID:    key.VectorStoreID,
			EmbeddingModelID: key.EmbeddingModelID,
			Type:             key.Type,
			KnowledgeIDs:     knowledgeIDs,
		})
	}
	return groups
}

// DeleteKnowledgeList deletes a knowledge entry and all related resources
func (s *knowledgeService) DeleteKnowledgeList(ctx context.Context, ids []string) error {
	plan, err := s.planKnowledgeDelete(ctx, ids)
	if err != nil {
		return err
	}
	return s.executeKnowledgeDelete(plan, false)
}

func (s *knowledgeService) executeKnowledgeDelete(plan *knowledgeDeletePlan, single bool) error {
	if len(plan.ids) == 0 {
		return nil
	}
	ctx, ids, knowledgeList := plan.ctx, plan.ids, plan.knowledge
	knowledgeBases, kbFileServices := plan.kbs, plan.files
	tenantInfo, _ := types.TenantInfoFromContext(ctx)

	// Mark all as deleting first to prevent async task conflicts.
	// Remember which entries still had queued / in-flight downstream tasks
	// so we can dequeue them in one pass after marking.
	var inFlightIDs []string
	for _, knowledge := range knowledgeList {
		prev := knowledge.ParseStatus
		before := *knowledge
		knowledge.ParseStatus = types.ParseStatusDeleting
		if !plan.rowsDeleted {
			if err := s.repo.UpdateKnowledgeForTransfer(ctx, &before, knowledge); err != nil {
				logger.GetLogger(ctx).WithField("error", err).WithField("knowledge_id", knowledge.ID).
					Errorf("DeleteKnowledgeList failed to mark as deleting")
				return err
			}
		}
		if prev == types.ParseStatusPending || prev == types.ParseStatusProcessing {
			inFlightIDs = append(inFlightIDs, knowledge.ID)
		}
	}
	logger.Infof(ctx, "Marked %d knowledge entries as deleting", len(knowledgeList))

	// Best-effort dequeue of downstream tasks for in-flight entries.
	// Workers also check deletion state; this avoids waking them unnecessarily.
	// The loop is per-knowledge because
	// the inspector only filters by knowledge_id, not by ID set.
	for _, kid := range inFlightIDs {
		s.dequeueKnowledgeTasks(ctx, kid)
	}

	chunkImageInfos := plan.imageInfo
	knowledgeToKB := make(map[string]string)
	for _, k := range knowledgeList {
		knowledgeToKB[k.ID] = k.KnowledgeBaseID
	}
	kbImageInfos := make(map[string][]string) // kbID → []imageInfo JSON
	for _, ci := range chunkImageInfos {
		kbID := knowledgeToKB[ci.KnowledgeID]
		kbImageInfos[kbID] = append(kbImageInfos[kbID], ci.ImageInfo)
	}
	kbImageURLs := make(map[string][]string) // kbID → []imageURL (deduplicated)
	for kbID, infos := range kbImageInfos {
		kbImageURLs[kbID] = collectImageURLs(ctx, infos)
	}
	kbKnowledgeIDs := make(map[string][]string) // kbID → knowledge IDs releasing their claims
	for _, k := range knowledgeList {
		kbKnowledgeIDs[k.KnowledgeBaseID] = append(kbKnowledgeIDs[k.KnowledgeBaseID], k.ID)
	}
	for kbID, ids := range kbKnowledgeIDs {
		kbImageURLs[kbID] = mergeKnowledgeReleaseURLs(ctx, s.resourceCatalog, ids, kbImageURLs[kbID])
	}

	wg := errgroup.Group{}
	// 2. Delete knowledge embeddings from vector store
	wg.Go(func() error {
		tenantID := types.MustTenantIDFromContext(ctx)
		for _, group := range buildKnowledgeVectorDeleteGroups(knowledgeList, knowledgeBases) {
			// Wiki-only knowledge never had embeddings written to the vector store,
			// and its EmbeddingModelID is intentionally empty. Skip the whole group
			// to avoid the spurious "model ID cannot be empty" failure.
			if strings.TrimSpace(group.EmbeddingModelID) == "" {
				logger.Infof(ctx, "Skipping vector store cleanup for %d knowledge entries without embedding model", len(group.KnowledgeIDs))
				continue
			}

			var vectorStoreID *string
			if group.VectorStoreID != "" {
				storeID := group.VectorStoreID
				vectorStoreID = &storeID
			}
			retrieveEngine, err := retriever.CreateRetrieveEngineForKB(
				ctx, s.retrieveEngine, s.ownership, tenantID, vectorStoreID)
			if err != nil {
				logger.GetLogger(ctx).WithField("error", err).Errorf("DeleteKnowledge delete knowledge embedding failed")
				return err
			}
			embeddingModel, err := s.modelService.GetEmbeddingModel(ctx, group.EmbeddingModelID)
			if err != nil {
				logger.GetLogger(ctx).WithField("error", err).Errorf("DeleteKnowledge get embedding model failed")
				return err
			}
			if err := retrieveEngine.DeleteByKnowledgeIDList(ctx, group.KnowledgeIDs, embeddingModel.GetDimensions(), group.Type); err != nil {
				logger.GetLogger(ctx).
					WithField("error", err).
					Errorf("DeleteKnowledge delete knowledge embedding failed")
				return err
			}
		}
		return nil
	})

	// 3. Clean wiki pages before deleting chunks so cleanup can still identify
	// which chunk_refs belonged to each source document.
	for _, knowledge := range knowledgeList {
		kb := knowledgeBases[knowledge.KnowledgeBaseID]
		if kb != nil && kb.IsWikiEnabled() {
			s.cleanupWikiOnKnowledgeDelete(ctx, knowledge)
		}
	}

	// 4. Delete all chunks associated with this knowledge
	wg.Go(func() error {
		if err := s.chunkRepo.DeleteByKnowledgeList(ctx, tenantInfo.ID, ids); err != nil {
			logger.GetLogger(ctx).WithField("error", err).Errorf("DeleteKnowledge delete chunks failed")
			return err
		}
		return nil
	})

	// Delete the knowledge graph
	wg.Go(func() error {
		namespaces := []types.NameSpace{}
		for _, knowledge := range knowledgeList {
			namespaces = append(
				namespaces,
				types.NameSpace{KnowledgeBase: knowledge.KnowledgeBaseID, Knowledge: knowledge.ID},
			)
		}
		if err := s.graphEngine.DelGraph(ctx, namespaces); err != nil {
			logger.GetLogger(ctx).WithField("error", err).Errorf("DeleteKnowledge delete knowledge graph failed")
			return err
		}
		return nil
	})

	if err := wg.Wait(); err != nil {
		return err
	}
	for _, knowledgeID := range ids {
		if err := s.repo.DeleteKnowledgeTagRelations(ctx, knowledgeID); err != nil {
			logger.Warnf(ctx, "Failed to delete tag relations for knowledge %s: %v", knowledgeID, err)
		}
	}
	// 6. Delete the knowledge rows FIRST, then drop their physical files. See
	// Deferring file removal until the rows are
	// gone avoids "file missing but row present" zombies that break reparse /
	// re-delete when an earlier cleanup step failed (issue #2192). A failure below
	// only orphans storage.
	if !plan.rowsDeleted {
		if err := s.repo.DeleteKnowledgeList(ctx, tenantInfo.ID, ids); err != nil {
			return err
		}
	}

	storageAdjust := int64(0)
	for _, knowledge := range knowledgeList {
		if knowledge.FilePath != "" {
			fSvc := kbFileServices[knowledge.KnowledgeBaseID]
			if err := fSvc.DeleteFile(ctx, knowledge.FilePath); err != nil {
				logger.GetLogger(ctx).WithField("error", err).Errorf("DeleteKnowledge delete file failed")
			}
		}
		storageAdjust -= knowledge.StorageSize
	}
	// Delete extracted images per KB
	for kbID, urls := range kbImageURLs {
		fSvc := kbFileServices[kbID]
		if fSvc == nil {
			logger.Warnf(ctx, "No file service for KB %s, skipping %d image deletions", kbID, len(urls))
			continue
		}
		deleteExtractedImages(ctx, fSvc, knowledgeResourceOwners(s.resourceCatalog, kbKnowledgeIDs[kbID]...), urls)
	}
	// TenantInfo can be shared by concurrent cleanup branches; update only storage accounting in the repository.
	if err := s.tenantRepo.AdjustStorageUsed(ctx, tenantInfo.ID, storageAdjust); err != nil {
		logger.GetLogger(ctx).WithField("error", err).Errorf("DeleteKnowledge update tenant storage used failed")
	}
	byKB := make(map[string][]*types.Knowledge)
	for i := range knowledgeList {
		knowledge := knowledgeList[i]
		byKB[knowledge.KnowledgeBaseID] = append(byKB[knowledge.KnowledgeBaseID], knowledge)
	}
	for kbID, knowledges := range byKB {
		// Deleted documents simply drop out of the description aggregation;
		// the refresh recomputes counts and topics from what remains.
		_ = requestKnowledgeBaseProfileRefresh(ctx, s.task, knowledgeBases[kbID], false)
		knowledgeIDs := make([]string, 0, len(knowledges))
		titles := make([]string, 0, len(knowledges))
		for _, knowledge := range knowledges {
			knowledgeIDs = append(knowledgeIDs, knowledge.ID)
			titles = append(titles, knowledge.Title)
		}
		details := map[string]any{"count": len(knowledgeIDs)}
		if len(knowledgeIDs) <= 20 {
			details["knowledge_ids"] = knowledgeIDs
		}
		kbActivityAppendSampleTitles(details, titles...)
		if single {
			knowledge := knowledges[0]
			recordKBActivity(ctx, s.audit, tenantInfo.ID, kbID, types.AuditActionKnowledgeDeleted,
				"knowledge", knowledge.ID, types.AuditOutcomeSuccess,
				map[string]any{"title": knowledge.Title, "type": knowledge.Type})
			continue
		}
		recordKBActivity(ctx, s.audit, tenantInfo.ID, kbID, types.AuditActionKnowledgeBatchDeleted,
			"knowledge", "", types.AuditOutcomeSuccess, details)
	}
	return nil
}

func (s *knowledgeService) cleanupKnowledgeResources(ctx context.Context, knowledge *types.Knowledge) error {
	logger.GetLogger(ctx).Infof("Cleaning knowledge resources before manual update, knowledge ID: %s", knowledge.ID)

	var cleanupErr error

	if knowledge.ParseStatus == types.ManualKnowledgeStatusDraft && knowledge.StorageSize == 0 {
		// Draft without indexed data, skip cleanup.
		return nil
	}

	kb, err := knowledgeWriteKB(ctx, s.kbService, knowledge)
	if err != nil {
		return fmt.Errorf("resolve cleanup knowledge base: %w", err)
	}
	tenantInfo := ctx.Value(types.TenantInfoContextKey).(*types.Tenant)
	if knowledge.EmbeddingModelID != "" {
		retrieveEngine, err := retriever.CreateRetrieveEngineForKB(
			ctx, s.retrieveEngine, s.ownership, tenantInfo.ID, kb.VectorStoreID)
		if err != nil {
			logger.GetLogger(ctx).WithField("error", err).Error("Failed to init retrieve engine during cleanup")
			cleanupErr = errors.Join(cleanupErr, err)
		} else {
			embeddingModel, modelErr := s.modelService.GetEmbeddingModel(ctx, knowledge.EmbeddingModelID)
			if modelErr != nil {
				logger.GetLogger(ctx).WithField("error", modelErr).Error("Failed to get embedding model during cleanup")
				cleanupErr = errors.Join(cleanupErr, modelErr)
			} else {
				if err := retrieveEngine.DeleteByKnowledgeIDList(ctx, []string{knowledge.ID}, embeddingModel.GetDimensions(), knowledge.Type); err != nil {
					logger.GetLogger(ctx).WithField("error", err).Error("Failed to delete manual knowledge index")
					cleanupErr = errors.Join(cleanupErr, err)
				}
			}
		}
	}

	// Collect image URLs before chunks are deleted
	fileSvc := s.resolveFileService(ctx, kb)
	chunkImageInfos, imgErr := s.chunkService.GetRepository().ListImageInfoByKnowledgeIDs(ctx, tenantInfo.ID, []string{knowledge.ID})
	if imgErr != nil {
		logger.GetLogger(ctx).WithField("error", imgErr).Error("Failed to collect image URLs for cleanup")
		cleanupErr = errors.Join(cleanupErr, imgErr)
	}
	var imageInfoStrs []string
	for _, ci := range chunkImageInfos {
		imageInfoStrs = append(imageInfoStrs, ci.ImageInfo)
	}
	// Release only what ImageInfo accounts for. This cleanup runs before the
	// knowledge is re-indexed (manual update, reparse, move with reparse), not
	// before it goes away, and the re-claim in triggerManualProcessing can only
	// succeed while the object is alive: DeleteFile marks the resource deleted,
	// and Bind resolves through a state=active lookup, so the handle could never
	// be re-bound. Unioning in catalog bindings here would therefore delete
	// markdown-only images that the new body still references. The paths that
	// actually remove a knowledge entry do union -- see mergeKnowledgeReleaseURLs.
	imageURLs := collectImageURLs(ctx, imageInfoStrs)

	if err := s.chunkRepo.DeleteChunksByKnowledgeID(ctx, knowledge.TenantID, knowledge.ID); err != nil {
		logger.GetLogger(ctx).WithField("error", err).Error("Failed to delete manual knowledge chunks")
		cleanupErr = errors.Join(cleanupErr, err)
	}

	// Delete extracted images after chunks are deleted. The claims released
	// here are re-taken by triggerManualProcessing, which always runs after
	// this cleanup and re-binds whatever the new body still references.
	deleteExtractedImages(ctx, fileSvc, knowledgeResourceOwners(s.resourceCatalog, knowledge.ID), imageURLs)

	namespace := types.NameSpace{KnowledgeBase: knowledge.KnowledgeBaseID, Knowledge: knowledge.ID}
	if err := s.graphEngine.DelGraph(ctx, []types.NameSpace{namespace}); err != nil {
		logger.GetLogger(ctx).WithField("error", err).Error("Failed to delete manual knowledge graph data")
		cleanupErr = errors.Join(cleanupErr, err)
	}

	if knowledge.StorageSize > 0 {
		tenantInfo.StorageUsed -= knowledge.StorageSize
		if tenantInfo.StorageUsed < 0 {
			tenantInfo.StorageUsed = 0
		}
		if err := s.tenantRepo.AdjustStorageUsed(ctx, tenantInfo.ID, -knowledge.StorageSize); err != nil {
			logger.GetLogger(ctx).WithField("error", err).Error("Failed to adjust storage usage during manual cleanup")
			cleanupErr = errors.Join(cleanupErr, err)
		}
		knowledge.StorageSize = 0
	}

	return cleanupErr
}

// ProcessKnowledgeListDelete handles Asynq knowledge list delete tasks
func (s *knowledgeService) ProcessKnowledgeListDelete(ctx context.Context, t *asynq.Task) error {
	var payload types.KnowledgeListDeletePayload
	if err := json.Unmarshal(t.Payload(), &payload); err != nil {
		logger.Errorf(ctx, "Failed to unmarshal knowledge list delete payload: %v", err)
		return err
	}
	ctx = payload.Initiator.Apply(ctx)
	taskID, _ := asynq.GetTaskID(ctx)
	ctx = withKBActivityTask(ctx, taskID, kbActivityTrigger(ctx))

	logger.Infof(ctx, "Processing knowledge list delete task for %d knowledge items", len(payload.KnowledgeIDs))

	if payload.TenantID == 0 {
		return fmt.Errorf("invalid delete task tenant: %w", asynq.SkipRetry)
	}
	ids, err := writeResourceIDs(payload.KnowledgeIDs)
	if err != nil {
		return fmt.Errorf("invalid delete task IDs: %w", asynq.SkipRetry)
	}
	if len(ids) == 0 {
		return nil
	}
	ctx = types.WithExecutionTenant(ctx, payload.TenantID)
	kbID := payload.KnowledgeBaseID
	if kbID == "" {
		// Pre-upgrade tasks were admitted by single-KB endpoints but omitted
		// the KB ID. Reconstruct only an unambiguous, same-tenant scope.
		rows, err := s.repo.GetKnowledgeBatch(ctx, payload.TenantID, ids)
		if err != nil {
			return err
		}
		if len(rows) == 0 {
			return nil
		}
		for _, row := range rows {
			if row == nil || row.TenantID != payload.TenantID || row.KnowledgeBaseID == "" ||
				(kbID != "" && row.KnowledgeBaseID != kbID) {
				return fmt.Errorf("ambiguous legacy delete scope: %w", asynq.SkipRetry)
			}
			kbID = row.KnowledgeBaseID
		}
	}
	bindings := make(map[string]string, len(ids))
	for _, id := range ids {
		bindings[id] = kbID
	}
	ctx = withKnowledgeCleanup(ctx, payload.TenantID, bindings)
	if err := s.DeleteKnowledgeList(ctx, ids); err != nil {
		var appErr *apperrors.AppError
		if errors.As(err, &appErr) && (appErr.HTTPCode == 403 || appErr.HTTPCode == 400 || appErr.HTTPCode == 409) {
			return fmt.Errorf("invalid delete task scope: %v: %w", err, asynq.SkipRetry)
		}
		return err
	}

	logger.Infof(ctx, "Successfully deleted %d knowledge items", len(payload.KnowledgeIDs))
	return nil
}