Skip to content

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

fix(retrieval): 分段中删除的图片不再被检索到

lawrence28dd15ec30d8
diff --git a/internal/application/repository/chunk.go b/internal/application/repository/chunk.go
index 17be8e557..ccb9fe871 100644
--- a/internal/application/repository/chunk.go
+++ b/internal/application/repository/chunk.go
@@ -160,6 +160,26 @@ func (r *chunkRepository) ListChunksByKnowledgeID(
return chunks, nil
}
+// ListChunksByKnowledgeIDAndTypes lists a knowledge's chunks restricted to the
+// given chunk types. ListChunksByKnowledgeID is text-only by design, so callers
+// that also need summary / parent_text / image chunks come through here rather
+// than widening that query underneath its existing callers.
+func (r *chunkRepository) ListChunksByKnowledgeIDAndTypes(
+ ctx context.Context, tenantID uint64, knowledgeID string, chunkTypes []types.ChunkType,
+) ([]*types.Chunk, error) {
+ if len(chunkTypes) == 0 {
+ return nil, nil
+ }
+ var chunks []*types.Chunk
+ if err := r.db.WithContext(ctx).
+ Where("tenant_id = ? AND knowledge_id = ? AND chunk_type IN ?", tenantID, knowledgeID, chunkTypes).
+ Order("chunk_index ASC").
+ Find(&chunks).Error; err != nil {
+ return nil, err
+ }
+ return chunks, nil
+}
+
// ListPagedChunksByKnowledgeID lists chunks for a knowledge ID with pagination
func (r *chunkRepository) ListPagedChunksByKnowledgeID(
ctx context.Context,
diff --git a/internal/application/service/chat_pipeline/merge_parent_resolve_test.go b/internal/application/service/chat_pipeline/merge_parent_resolve_test.go
index 67516a3ff..5ae6688d7 100644
--- a/internal/application/service/chat_pipeline/merge_parent_resolve_test.go
+++ b/internal/application/service/chat_pipeline/merge_parent_resolve_test.go
@@ -26,6 +26,7 @@ func TestResolveParentChunksUsesCurrentContentAndImageURLs(t *testing.T) {
{
ID: "image", ParentChunkID: "child", ChunkType: types.ChunkTypeImageOCR,
ImageInfo: string(imageInfo),
+ IsEnabled: true,
},
},
},
@@ -73,6 +74,7 @@ func TestResolveImageChunkKeepsGrandparentContextWithoutCoordinateSlicing(t *tes
"text": {{
ID: "image", ParentChunkID: "text", ChunkType: types.ChunkTypeImageOCR,
ImageInfo: string(imageInfo),
+ IsEnabled: true,
}},
},
}
diff --git a/internal/application/service/knowledge_process.go b/internal/application/service/knowledge_process.go
index 3ac257843..6c2f7350b 100644
--- a/internal/application/service/knowledge_process.go
+++ b/internal/application/service/knowledge_process.go
@@ -2340,26 +2340,35 @@ func (s *knowledgeService) RegenerateKnowledgeSummary(
return nil, err
}
if kb.NeedsEmbeddingModel() {
- found := false
maxIndex := 0
- summaryChunks := make([]*types.Chunk, 0, 1)
for _, chunk := range allChunks {
if chunk.ChunkIndex > maxIndex {
maxIndex = chunk.ChunkIndex
}
- if chunk.ChunkType == types.ChunkTypeSummary {
- chunk.Content = "# Summary\n" + summary
- chunk.SourceContent = chunk.Content
- chunk.IsEnabled = true
- chunk.UpdatedAt = time.Now()
- if err := s.chunkRepo.UpdateChunk(ctx, chunk); err != nil {
- return nil, err
- }
- summaryChunks = append(summaryChunks, chunk)
- found = true
+ }
+ // allChunks holds text chunks only, so it can never carry the existing
+ // summary chunk. Scanning it for one always came up empty, which left
+ // every refresh appending a new summary chunk beside the stale one --
+ // and a stale summary stays enabled and indexed, so content the user
+ // edited out of the document kept being retrievable through it.
+ existingSummaries, err := s.chunkRepo.ListChunksByKnowledgeIDAndTypes(
+ ctx, tenantID, knowledgeID, []types.ChunkType{types.ChunkTypeSummary},
+ )
+ if err != nil {
+ return nil, err
+ }
+ summaryChunks := make([]*types.Chunk, 0, len(existingSummaries))
+ for _, chunk := range existingSummaries {
+ chunk.Content = "# Summary\n" + summary
+ chunk.SourceContent = chunk.Content
+ chunk.IsEnabled = true
+ chunk.UpdatedAt = time.Now()
+ if err := s.chunkRepo.UpdateChunk(ctx, chunk); err != nil {
+ return nil, err
}
+ summaryChunks = append(summaryChunks, chunk)
}
- if !found {
+ if len(summaryChunks) == 0 {
summaryChunk := &types.Chunk{
ID: uuid.NewString(), TenantID: tenantID, KnowledgeID: knowledge.ID,
KnowledgeBaseID: knowledge.KnowledgeBaseID, Content: "# Summary\n" + summary,
@@ -2973,6 +2982,10 @@ func (s *knowledgeService) UpdateImageInfo(
ChunkType: types.ChunkTypeImageCaption,
ParentChunkID: chunk.ID,
ImageInfo: imageInfo,
+ // CreateChunks inserts with Select("*"), so the gorm default:true never
+ // applies -- an unset IsEnabled lands in the database as false and the
+ // chunk is silently excluded from retrieval and model context.
+ IsEnabled: true,
}
addChunk = append(addChunk, captionChunk)
logger.Infof(ctx, "Created new caption chunk ID: %s for image URL: %s", captionChunk.ID, image.OriginalURL)
@@ -2989,6 +3002,7 @@ func (s *knowledgeService) UpdateImageInfo(
ChunkType: types.ChunkTypeImageOCR,
ParentChunkID: chunk.ID,
ImageInfo: imageInfo,
+ IsEnabled: true,
}
addChunk = append(addChunk, ocrChunk)
logger.Infof(ctx, "Created new OCR chunk ID: %s for image URL: %s", ocrChunk.ID, image.OriginalURL)
diff --git a/internal/searchutil/imageinfo.go b/internal/searchutil/imageinfo.go
index 7cba563fd..487563aea 100644
--- a/internal/searchutil/imageinfo.go
+++ b/internal/searchutil/imageinfo.go
@@ -46,6 +46,11 @@ const HTMLImageSrcURLGroup = 2
// - If chunkIDs are parent_text chunks, their children are text chunks
// whose children are image chunks → two queries.
//
+// Disabled children are skipped at both levels. Removing an image from a chunk
+// disables its image_ocr / image_caption children (syncEditedChunkImages), and
+// disabling a text chunk disables them too; honoring that flag here is what
+// keeps a deleted image out of retrieval, summaries and model context.
+//
// Returns a map of input chunkID → merged image_info JSON string.
func CollectImageInfoByChunkIDs(
ctx context.Context,
@@ -107,6 +112,9 @@ func CollectImageInfoByChunkIDs(
textToParent := make(map[string]string)
for _, child := range children {
+ if !child.IsEnabled {
+ continue
+ }
switch child.ChunkType {
case types.ChunkTypeImageOCR, types.ChunkTypeImageCaption:
addInfo(child.ParentChunkID, child)
@@ -120,6 +128,9 @@ func CollectImageInfoByChunkIDs(
grandChildren, err := chunkRepo.ListChunksByParentIDs(ctx, tenantID, textChildIDs)
if err == nil {
for _, gc := range grandChildren {
+ if !gc.IsEnabled {
+ continue
+ }
if gc.ChunkType != types.ChunkTypeImageOCR && gc.ChunkType != types.ChunkTypeImageCaption {
continue
}
diff --git a/internal/searchutil/imageinfo_disabled_test.go b/internal/searchutil/imageinfo_disabled_test.go
new file mode 100644
index 000000000..6d4e41495
--- /dev/null
+++ b/internal/searchutil/imageinfo_disabled_test.go
@@ -0,0 +1,84 @@
+package searchutil
+
+import (
+ "context"
+ "testing"
+
+ "github.com/Tencent/WeKnora/internal/types"
+ "github.com/Tencent/WeKnora/internal/types/interfaces"
+)
+
+// childrenByParentRepo serves chunk children from a static parent → children map.
+type childrenByParentRepo struct {
+ interfaces.ChunkRepository
+ byParent map[string][]*types.Chunk
+}
+
+func (r childrenByParentRepo) ListChunksByParentIDs(
+ _ context.Context, _ uint64, parentIDs []string,
+) ([]*types.Chunk, error) {
+ var out []*types.Chunk
+ for _, id := range parentIDs {
+ out = append(out, r.byParent[id]...)
+ }
+ return out, nil
+}
+
+func imageChild(id, parentID string, chunkType types.ChunkType, url string, enabled bool) *types.Chunk {
+ return &types.Chunk{
+ ID: id,
+ ParentChunkID: parentID,
+ ChunkType: chunkType,
+ IsEnabled: enabled,
+ ImageInfo: `[{"url":"` + url + `","caption":"a screenshot"}]`,
+ }
+}
+
+// Deleting an image from a chunk disables its image children. Their image_info
+// must not come back through enrichment, or the removed image reappears in
+// summaries, tool output and model context.
+func TestCollectImageInfoByChunkIDsSkipsDisabledImageChildren(t *testing.T) {
+ repo := childrenByParentRepo{byParent: map[string][]*types.Chunk{
+ "text-1": {
+ imageChild("ocr-1", "text-1", types.ChunkTypeImageOCR, "resource://removed", false),
+ imageChild("cap-1", "text-1", types.ChunkTypeImageCaption, "resource://removed", false),
+ },
+ "text-2": {
+ imageChild("ocr-2", "text-2", types.ChunkTypeImageOCR, "resource://kept", true),
+ },
+ }}
+
+ got := CollectImageInfoByChunkIDs(context.Background(), repo, 1, []string{"text-1", "text-2"})
+
+ if _, ok := got["text-1"]; ok {
+ t.Fatalf("disabled image children leaked image_info: %q", got["text-1"])
+ }
+ if got["text-2"] == "" {
+ t.Fatal("enabled image child lost its image_info")
+ }
+}
+
+// The parent_text → text → image path must apply the same rule at both levels.
+func TestCollectImageInfoByChunkIDsSkipsDisabledGrandchildrenAndTextChildren(t *testing.T) {
+ repo := childrenByParentRepo{byParent: map[string][]*types.Chunk{
+ "parent-1": {
+ {ID: "text-1", ParentChunkID: "parent-1", ChunkType: types.ChunkTypeText, IsEnabled: true},
+ },
+ "parent-2": {
+ // Disabling a text chunk hides everything under it.
+ {ID: "text-2", ParentChunkID: "parent-2", ChunkType: types.ChunkTypeText, IsEnabled: false},
+ },
+ "text-1": {
+ imageChild("ocr-1", "text-1", types.ChunkTypeImageOCR, "resource://removed", false),
+ },
+ "text-2": {
+ imageChild("ocr-2", "text-2", types.ChunkTypeImageOCR, "resource://hidden", true),
+ },
+ }}
+
+ got := CollectImageInfoByChunkIDs(context.Background(), repo, 1, []string{"parent-1", "parent-2"})
+
+ if len(got) != 0 {
+ t.Fatalf("expected no image_info, got %v", got)
+ }
+}
diff --git a/internal/types/interfaces/chunk.go b/internal/types/interfaces/chunk.go
index 9f6434d61..28cf8a4c6 100644
--- a/internal/types/interfaces/chunk.go
+++ b/internal/types/interfaces/chunk.go
@@ -28,8 +28,15 @@ type ChunkRepository interface {
ListChunksByIDOnly(ctx context.Context, ids []string) ([]*types.Chunk, error)
// ListChunksBySeqID lists chunks by seq_ids
ListChunksBySeqID(ctx context.Context, tenantID uint64, seqIDs []int64) ([]*types.Chunk, error)
- // ListChunksByKnowledgeID lists chunks by knowledge id
+ // ListChunksByKnowledgeID lists the knowledge's text chunks. Despite the name
+ // it filters to chunk_type = 'text'; reach for ListChunksByKnowledgeIDAndTypes
+ // when summary, parent_text or image chunks are needed too.
ListChunksByKnowledgeID(ctx context.Context, tenantID uint64, knowledgeID string) ([]*types.Chunk, error)
+ // ListChunksByKnowledgeIDAndTypes lists the knowledge's chunks restricted to
+ // the given chunk types, ordered by chunk_index.
+ ListChunksByKnowledgeIDAndTypes(
+ ctx context.Context, tenantID uint64, knowledgeID string, chunkTypes []types.ChunkType,
+ ) ([]*types.Chunk, error)
// ListPagedChunksByKnowledgeID lists paged chunks by knowledge id.
// When tagIDs is non-empty, results are filtered by tag_id (OR semantics).
// knowledgeType: "faq" or "manual" - determines sort order and search behavior