Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

fix(deepdoc): preserve complete PDF figures (#21390)

jay77721fd675ebde21a
diff --git a/internal/deepdoc/parser/pdf/figure_bounds_integration_test.go b/internal/deepdoc/parser/pdf/figure_bounds_integration_test.go
new file mode 100644
index 0000000000..5273402cb7
--- /dev/null
+++ b/internal/deepdoc/parser/pdf/figure_bounds_integration_test.go
@@ -0,0 +1,54 @@
+//go:build cgo && integration
+
+package pdf
+
+import (
+ "fmt"
+ "os"
+ "testing"
+
+ pdf "ragflow/internal/deepdoc/parser/pdf/type"
+)
+
+// TestAN0590FigureBounds exercises complete DLA figure bounds on the real PDF.
+// AN0590_PDF points to an0590scv140.pdf; MODEL_DIR supplies DeepDOC models.
+func TestAN0590FigureBounds(t *testing.T) {
+ path := os.Getenv("AN0590_PDF")
+ if path == "" {
+ t.Skip("AN0590_PDF is required")
+ }
+ analyzer := mustConnectInProcessAnalyzer(t)
+ data, err := os.ReadFile(path)
+ if err != nil {
+ t.Fatal(err)
+ }
+ cfg := pdf.DefaultParserConfig()
+ result, err := NewParser(cfg).Parse(t.Context(), data, analyzer)
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer result.Close()
+ for _, tc := range []struct {
+ number, page int
+ minWidth, minHeight float64
+ }{
+ {2, 1, 240, 50}, {3, 3, 350, 100},
+ } {
+ t.Run(fmt.Sprintf("figure%d", tc.number), func(t *testing.T) {
+ var position *pdf.Position
+ for _, section := range result.Figures() {
+ for i := range section.Positions {
+ pos := &section.Positions[i]
+ if len(pos.PageNumbers) == 1 && pos.PageNumbers[0] == tc.page &&
+ pos.Right-pos.Left >= tc.minWidth && pos.Bottom-pos.Top >= tc.minHeight {
+ position = pos
+ break
+ }
+ }
+ }
+ if position == nil {
+ t.Fatalf("figure %d missing", tc.number)
+ }
+ })
+ }
+}
diff --git a/internal/deepdoc/parser/pdf/figure_bounds_test.go b/internal/deepdoc/parser/pdf/figure_bounds_test.go
new file mode 100644
index 0000000000..faaaa604ba
--- /dev/null
+++ b/internal/deepdoc/parser/pdf/figure_bounds_test.go
@@ -0,0 +1,46 @@
+package pdf
+
+import (
+ "image"
+ "testing"
+
+ tbl "ragflow/internal/deepdoc/parser/pdf/table"
+ pdf "ragflow/internal/deepdoc/parser/pdf/type"
+)
+
+func TestEnrichFigureBoundsUseRenderScale(t *testing.T) {
+ analyzer := &MockDocAnalyzer{Healthy: true, DLARegions: []pdf.DLARegion{
+ {X0: 90, Y0: 180, X1: 810, Y1: 540, Label: "figure", Confidence: 0.95},
+ }}
+ boxes := []pdf.TextBox{{Text: "SDA", X0: 5, X1: 25, Top: 30, Bottom: 35}}
+ p := NewParser(pdf.DefaultParserConfig())
+ annotated, _, _ := p.enrichOnePageWithDeepDoc(t.Context(), image.NewRGBA(image.Rect(0, 0, 900, 900)), boxes, 4, nil, analyzer, NewTableBuilderFor(analyzer), 9)
+ if len(annotated) != 2 || annotated[0].X0 != 5 || annotated[0].X1 != 25 {
+ t.Fatalf("annotation changed native text geometry: %+v", annotated)
+ }
+ figures := tbl.ConsolidateFigures(annotated)
+ if len(figures) != 1 || figures[0].PageNumber != 4 || figures[0].X0 != 10 || figures[0].X1 != 90 || figures[0].Top != 20 || figures[0].Bottom != 60 {
+ t.Fatalf("figure did not retain retry-render scale or page: %+v", figures)
+ }
+}
+
+func TestParsePreservesFigureWithoutText(t *testing.T) {
+ engine := &MockEngine{NumPages: 4, RenderW: 900, RenderH: 900}
+ analyzer := &MockDocAnalyzer{Healthy: true, DLARegions: []pdf.DLARegion{
+ {X0: 90, Y0: 180, X1: 810, Y1: 540, Label: "figure", Confidence: 0.95},
+ }}
+ cfg := pdf.DefaultParserConfig()
+ cfg.Pages = [][]int{{4, 4}}
+ result, err := NewParser(cfg).ParseRaw(t.Context(), engine, analyzer)
+ if err != nil {
+ t.Fatal(err)
+ }
+ figures := result.Figures()
+ if len(figures) != 1 || figures[0].Text != "" {
+ t.Fatalf("textless figure lost or fabricated text: %+v", result.Sections)
+ }
+ pos := figures[0].Positions[0]
+ if len(pos.PageNumbers) != 1 || pos.PageNumbers[0] != 3 || pos.Left != 30 || pos.Right != 270 || pos.Top != 60 || pos.Bottom != 180 {
+ t.Fatalf("textless figure geometry lost: %+v", pos)
+ }
+}
diff --git a/internal/deepdoc/parser/pdf/layout/boxes_sections.go b/internal/deepdoc/parser/pdf/layout/boxes_sections.go
index bfc5dc3082..cdbe865cc6 100644
--- a/internal/deepdoc/parser/pdf/layout/boxes_sections.go
+++ b/internal/deepdoc/parser/pdf/layout/boxes_sections.go
@@ -55,7 +55,7 @@ func BoxesToSections(boxes []pdf.TextBox, pageHeights map[int]float64) []pdf.Sec
sections := make([]pdf.Section, 0, len(boxes))
for _, b := range boxes {
t := strings.TrimSpace(b.Text)
- if t == "" {
+ if t == "" && b.LayoutType != pdf.LayoutTypeFigure {
continue
}
diff --git a/internal/deepdoc/parser/pdf/layout/layout.go b/internal/deepdoc/parser/pdf/layout/layout.go
index 9708d228bc..79946f5cd2 100644
--- a/internal/deepdoc/parser/pdf/layout/layout.go
+++ b/internal/deepdoc/parser/pdf/layout/layout.go
@@ -554,6 +554,9 @@ func groupBoxesByPage(boxes []pdf.TextBox) (map[int][]int, []int) {
// shouldMergeBoxes determines whether two boxes should be merged
func shouldMergeBoxes(prev, curr *pdf.TextBox, mh, mw float64, isEnglish bool) bool {
+ if prev.LayoutType == pdf.LayoutTypeFigure || curr.LayoutType == pdf.LayoutTypeFigure {
+ return false
+ }
// Check layout number
if prev.LayoutNo != curr.LayoutNo {
common.Debug("vm reject", zap.String("reason", "layoutNo"),
@@ -674,6 +677,10 @@ func mergeColumnBoxes(sortedBoxes []pdf.TextBox, mh, mw float64, isEnglish bool)
out := make([]pdf.TextBox, 0, len(sortedBoxes))
for i := 0; i < len(sortedBoxes); i++ {
curr := sortedBoxes[i]
+ if curr.LayoutType == pdf.LayoutTypeFigure {
+ out = append(out, curr)
+ continue
+ }
// Skip cross-page suffixes (like previous page number)
if i > 0 && sortedBoxes[i-1].PageNumber < curr.PageNumber && pageNumSuffixPattern.MatchString(sortedBoxes[i-1].Text) {
diff --git a/internal/deepdoc/parser/pdf/table/merge_captions.go b/internal/deepdoc/parser/pdf/table/merge_captions.go
index eb3ead582e..f4264aa8b9 100644
--- a/internal/deepdoc/parser/pdf/table/merge_captions.go
+++ b/internal/deepdoc/parser/pdf/table/merge_captions.go
@@ -84,16 +84,13 @@ func MergeCaptions(sections []pdf.Section) []pdf.Section {
captions = append(captions, i)
continue
}
- // No merge target. A FIGURE caption is kept as its own section: a pure
- // image figure has no text section (BoxesToSections skips empty figure
- // boxes), so removing it would drop caption text that Python keeps
- // (07_mixed_content 'Figure 1/2'). A TABLE caption with neither a table
- // nor a figure parent — findNearestParent already fell back to figures,
- // mirroring Python's nearest(tables)/nearest(figures) double search — is
- // a DLA mislabel (e.g. rotate_270's rotated text labeled "table"); keep
- // the historical removal so rotated-page text is not duplicated. Python
- // likewise drops a caption only when BOTH searches fail, so the removal
- // branch now fires only when no parent exists at all.
+ // No merge target. Keep a FIGURE caption as its own section when no
+ // nearby figure is available. Position-only sections let captions for
+ // pure-image figures attach to their image before reaching this branch.
+ // A TABLE caption with no reachable table or figure parent is a DLA
+ // mislabel (e.g. rotate_270's rotated text labeled "table").
+ // findNearestParent searches both kinds, matching Python's fallback.
+ // Keep the historical removal so rotated-page text is not duplicated.
if captionType != pdf.LayoutTypeFigure {
captions = append(captions, i)
}
diff --git a/internal/deepdoc/parser/pdf/table/table_annotate.go b/internal/deepdoc/parser/pdf/table/table_annotate.go
index 73770b4712..dcb50998fe 100644
--- a/internal/deepdoc/parser/pdf/table/table_annotate.go
+++ b/internal/deepdoc/parser/pdf/table/table_annotate.go
@@ -37,6 +37,12 @@ func regionOverlapsBox(region pdf.DLARegion, box pdf.TextBox, scale float64) boo
// Each table region is matched if at least one box overlaps it (>40% of box
// area) or if there are no boxes at all (image-only PDF), matching Python's
// _table_transformer_job which processes every table DLA region.
+//
+// Figure-layout boxes are skipped: AnnotateBoxLayouts appends one full-extent
+// geometry box per figure/equation region, and a nested/double-detected
+// table+figure pair would otherwise pull the whole figure extent into the
+// table's BoxIdx (and from there into Positions, boxInCrop and R/C
+// annotations, causing HTML cross-region and NoMerge mislabels).
func MatchTableRegions(boxes []pdf.TextBox, regions []pdf.DLARegion, scale float64) []TableMatch {
var matches []TableMatch
for _, r := range regions {
@@ -45,6 +51,9 @@ func MatchTableRegions(boxes []pdf.TextBox, regions []pdf.DLARegion, scale float
}
var matched []int
for i, b := range boxes {
+ if b.LayoutType == pdf.LayoutTypeFigure {
+ continue
+ }
if regionOverlapsBox(r, b, scale) {
matched = append(matched, i)
}
@@ -65,7 +74,6 @@ type annRegion struct {
x0, y0, x1, y1 float64
label string
score float64
- visited bool
typeIndex int
}
@@ -280,11 +288,11 @@ func nmsIoU(a, b pdf.DLARegion) float64 {
// AnnotateBoxLayouts assigns LayoutType and LayoutNo to boxes based on DLA
// regions. Returns the filtered slice (Python pops CID-garbled boxes and
// garbage-layout boxes at wrong positions - Go mirrors with compact).
-// Also creates synthetic figure boxes for unmatched figure/equation regions.
+// Also creates a complete geometry box for every figure/equation region.
//
// Before annotation, regions are de-duplicated (layouts_cleanup) and sorted
-// top-to-bottom (sort_Y_firstly) to match Python, and unmatched figure and
-// equation regions receive SEPARATE synthetic namespaces (figure-N /
+// top-to-bottom (sort_Y_firstly), and figure and equation regions receive
+// separate geometry namespaces (figure-N /
// equation-N) so they never collide.
// FilteredDLARegions returns the DLA regions after per-class NMS, the
// confidence filter, Y-sort, and cleanup — i.e. the exact set fed to
@@ -413,7 +421,6 @@ func AnnotateBoxLayouts(boxes []pdf.TextBox, regions []pdf.DLARegion, scale floa
dropped[i] = true
continue
}
- cands[bestJ].visited = true
// Python: equation mapped to "figure" for layout_type
if ty == pdf.LayoutTypeEquation {
boxes[i].LayoutType = pdf.LayoutTypeFigure
@@ -446,28 +453,21 @@ func AnnotateBoxLayouts(boxes []pdf.TextBox, regions []pdf.DLARegion, scale floa
}
boxes = compacted
- // Synthetic figure boxes for unmatched figure/equation regions (Python:
- // layout_recognizer.py:145-155). Python numbers each unmatched region with
- // its index WITHIN the per-type list that also includes already-visited
- // regions (enumerate([lt for lt in lts if lt["type"] == ty])), so we reuse
- // the per-type typeIndex computed above rather than a separate
- // unvisited-only counter. Python keeps figure-N / equation-N in SEPARATE
- // namespaces, so the typeIndex is keyed by the original type label.
+ // Keep one geometry box per figure/equation, independently of its text.
+ // ConsolidateFigures combines it with matching labels after text cleanup.
for j := range cands {
- if cands[j].visited {
- continue
- }
if cands[j].label != pdf.LayoutTypeFigure && cands[j].label != pdf.LayoutTypeEquation {
continue
}
boxes = append(boxes, pdf.TextBox{
- X0: cands[j].x0,
- X1: cands[j].x1,
- Top: cands[j].y0,
- Bottom: cands[j].y1,
- Text: "",
- LayoutType: pdf.LayoutTypeFigure,
- LayoutNo: fmt.Sprintf("%s-%d", cands[j].label, cands[j].typeIndex),
+ X0: cands[j].x0,
+ X1: cands[j].x1,
+ Top: cands[j].y0,
+ Bottom: cands[j].y1,
+ Text: "",
+ LayoutType: pdf.LayoutTypeFigure,
+ LayoutNo: fmt.Sprintf("%s-%d", cands[j].label, cands[j].typeIndex),
+ FigureBounds: &pdf.Position{Left: cands[j].x0, Top: cands[j].y0, Right: cands[j].x1, Bottom: cands[j].y1},
})
}
diff --git a/internal/deepdoc/parser/pdf/table/table_annotate_parity_test.go b/internal/deepdoc/parser/pdf/table/table_annotate_parity_test.go
index 23cb815e1a..d1664c086e 100644
--- a/internal/deepdoc/parser/pdf/table/table_annotate_parity_test.go
+++ b/internal/deepdoc/parser/pdf/table/table_annotate_parity_test.go
@@ -1,24 +1,5 @@
-// Tests pinning the Python-equivalent behavior of AnnotateBoxLayouts.
-//
-// These assert the layout-annotation semantics that Python's
-// LayoutRecognizer.__call__ (deepdoc/vision/layout_recognizer.py:68) produces,
-// and that the Go implementation now replicates (GREEN):
-//
-// - #1 layouts_cleanup: Python de-dupes overlapping same-type regions
-// (recognizer.py:124-160, called at layout_recognizer.py:100) BEFORE
-// annotation. Go matches via cleanupLayouts, so no extra synthetic
-// figure/equation boxes are emitted.
-// - #2 sort_Y_firstly: Python sorts regions top-to-bottom (recognizer.py:54,
-// layout_recognizer.py:99) before numbering them, so layoutno indices
-// follow reading order. Go matches via sortYFirstly.
-// - #3 synthetic namespace: Python numbers unmatched figure/equation regions
-// in SEPARATE counters -> "figure-N" / "equation-N"
-// (layout_recognizer.py:145-155). Go matches via a per-type typeIndex
-// keyed by the original type label.
-//
-// These are pure-Go unit tests (no model server, no external service): they
-// feed crafted pdf.DLARegion slices and assert the annotated pdf.TextBox
-// output. Coordinates use scale=1 so region pixels == PDF space.
+// Tests for layout filtering, reading-order numbering, and figure identity.
+// Text annotations and complete image regions must share the same identifiers.
package table
@@ -40,8 +21,7 @@ func countFigureBoxes(boxes []pdf.TextBox) int {
}
// countSyntheticFigures returns the number of figure boxes with no text
-// (the synthetic placeholders AnnotateBoxLayouts appends for unvisited
-// figure/equation regions).
+// (the geometry boxes emitted for figure/equation regions).
func countSyntheticFigures(boxes []pdf.TextBox) int {
n := 0
for _, b := range boxes {
@@ -52,12 +32,7 @@ func countSyntheticFigures(boxes []pdf.TextBox) int {
return n
}
-// TestAnnotateBoxLayouts_DuplicateFigureRegions_Merged pins #1: two heavily
-// overlapping same-type (figure) regions must be treated as ONE, so only one
-// figure box (the annotated text box) results and NO synthetic figure box is
-// produced. Python merges via layouts_cleanup; Go now mirrors that by fusing
-// the overlapping pair into one region, so the duplicate synthetic figure is
-// avoided.
+// Duplicate DLA detections must assemble into one complete figure.
func TestAnnotateBoxLayouts_DuplicateFigureRegions_Merged(t *testing.T) {
box := pdf.TextBox{X0: 0, X1: 100, Top: 0, Bottom: 50, Text: "Some caption text", PageNumber: 0}
regions := []pdf.DLARegion{
@@ -65,7 +40,7 @@ func TestAnnotateBoxLayouts_DuplicateFigureRegions_Merged(t *testing.T) {
{X0: 0, Y0: 0, X1: 100, Y1: 50, Confidence: 0.8, Label: pdf.LayoutTypeFigure}, // identical -> IoU 1.0
}
- out := AnnotateBoxLayouts([]pdf.TextBox{box}, regions, 1.0, 100.0)
+ out := ConsolidateFigures(AnnotateBoxLayouts([]pdf.TextBox{box}, regions, 1.0, 100.0))
if got := countFigureBoxes(out); got != 1 {
t.Errorf("#1 layouts_cleanup: expected 1 figure box after de-duplicating overlapping figure regions, got %d", got)
@@ -242,7 +217,7 @@ func TestAnnotateBoxLayouts_CleanupEqualScoreKeepsLater(t *testing.T) {
{X0: 0, Y0: 0, X1: 400, Y1: 400, Confidence: 0.9, Label: pdf.LayoutTypeFigure}, // B: large, on top (Y0=0)
{X0: 100, Y0: 200, X1: 200, Y1: 300, Confidence: 0.9, Label: pdf.LayoutTypeFigure}, // A: small, inside B, lower (Y0=200)
}
- // No text box overlaps -> both unvisited -> cleanup collapses to one synthetic.
+ // Cleanup must retain only one geometry box for these overlapping regions.
out := AnnotateBoxLayouts([]pdf.TextBox{}, regions, 1.0, 100.0)
if got := countFigureBoxes(out); got != 1 {
t.Fatalf("expected 1 figure box after cleanup merge, got %d", got)
@@ -264,7 +239,7 @@ func TestAnnotateBoxLayouts_SyntheticVisitedInterleaved(t *testing.T) {
{X0: 0, X1: 100, Top: 0, Bottom: 50, Text: "caption for A"},
}
regions := []pdf.DLARegion{
- {X0: 0, Y0: 0, X1: 100, Y1: 50, Confidence: 0.9, Label: pdf.LayoutTypeFigure}, // A: matched to text box -> visited
+ {X0: 0, Y0: 0, X1: 100, Y1: 50, Confidence: 0.9, Label: pdf.LayoutTypeFigure}, // A: matched to text box -> annotated
{X0: 200, Y0: 0, X1: 300, Y1: 50, Confidence: 0.9, Label: pdf.LayoutTypeFigure}, // B: unmatched -> synthetic
}
out := AnnotateBoxLayouts(boxes, regions, 1.0, 100.0)
@@ -279,15 +254,15 @@ func TestAnnotateBoxLayouts_SyntheticVisitedInterleaved(t *testing.T) {
}
}
if annotated == nil {
- t.Fatal("visited figure A: text box was not annotated as figure")
+ t.Fatal("annotated figure A: text box was not annotated as figure")
}
if annotated.LayoutNo != "figure-0" {
- t.Errorf("visited figure A: expected LayoutNo figure-0, got %q", annotated.LayoutNo)
+ t.Errorf("annotated figure A: expected LayoutNo figure-0, got %q", annotated.LayoutNo)
}
if synthetic == nil {
t.Fatal("expected a synthetic figure box for unmatched B")
}
- // B is the 2nd figure in Y order (after visited A), so Python numbers it figure-1.
+ // B is the 2nd figure in Y order (after annotated A), so Python numbers it figure-1.
if synthetic.LayoutNo != "figure-1" {
t.Errorf("unmatched figure B: expected figure-1 (index in full per-type list), got %q", synthetic.LayoutNo)
}
diff --git a/internal/deepdoc/parser/pdf/table/table_annotate_test.go b/internal/deepdoc/parser/pdf/table/table_annotate_test.go
index 4138ac8140..0b440681a4 100644
--- a/internal/deepdoc/parser/pdf/table/table_annotate_test.go
+++ b/internal/deepdoc/parser/pdf/table/table_annotate_test.go
@@ -349,9 +349,9 @@ func TestAnnotateBoxLayouts_EquationMappedToFigure(t *testing.T) {
regions := []pdf.DLARegion{
{X0: 0, Y0: 0, X1: 300, Y1: 60, Label: "equation", Confidence: 0.9},
}
- boxes = AnnotateBoxLayouts(boxes, regions, 3.0, 0)
+ boxes = ConsolidateFigures(AnnotateBoxLayouts(boxes, regions, 3.0, 0))
if len(boxes) != 1 {
- t.Fatalf("expected 1 box, got %d", len(boxes))
+ t.Fatalf("unexpected assembled box count: %d", len(boxes))
}
if boxes[0].LayoutType != "figure" {
t.Errorf("equation → LayoutType: got %q, want 'figure'", boxes[0].LayoutType)
@@ -374,9 +374,9 @@ func TestAnnotateBoxLayouts_MixedTypesLayoutNo(t *testing.T) {
{X0: 0, Y0: 600, X1: 150, Y1: 660, Label: "text", Confidence: 0.9}, // text-1
{X0: 600, Y0: 0, X1: 900, Y1: 60, Label: "figure", Confidence: 0.9}, // figure-0 (PDF: x0=200, x1=300)
}
- boxes = AnnotateBoxLayouts(boxes, regions, 3.0, 0)
+ boxes = ConsolidateFigures(AnnotateBoxLayouts(boxes, regions, 3.0, 0))
if len(boxes) != 3 {
- t.Fatalf("expected 3 boxes, got %d", len(boxes))
+ t.Fatalf("unexpected assembled box count: %d", len(boxes))
}
// Check that text and figure indices are independent
if boxes[0].LayoutNo != "text-0" {
@@ -481,6 +481,26 @@ func TestMatchTableRegions_BelowThreshold(t *testing.T) {
}
}
+func TestMatchTableRegions_SkipsFigureBoxes(t *testing.T) {
+ // Nested/double-detected table+figure: the full-extent figure geometry
+ // box must not enter the table's BoxIdx, otherwise its extent leaks
+ // into TableItem.Positions (HTML cross-region, NoMerge mislabel).
+ boxes := []pdf.TextBox{
+ {X0: 0, X1: 100, Top: 0, Bottom: 50, LayoutType: pdf.LayoutTypeTable},
+ {X0: 0, X1: 100, Top: 0, Bottom: 50, LayoutType: pdf.LayoutTypeFigure},
+ }
+ regions := []pdf.DLARegion{
+ {X0: 0, Y0: 0, X1: 300, Y1: 150, Label: "table"},
+ }
+ matches := MatchTableRegions(boxes, regions, 3.0)
+ if len(matches) != 1 {
+ t.Fatalf("expected 1 match, got %d", len(matches))
+ }
+ if len(matches[0].BoxIdx) != 1 || matches[0].BoxIdx[0] != 0 {
+ t.Errorf("figure box leaked into table match, got %v", matches[0].BoxIdx)
+ }
+}
+
// MockTableBuilder is a test-only pdf.TableBuilder with a configurable GroupCells.
type MockTableBuilder struct {
GroupCellsFn func(cells []pdf.TSRCell) [][]pdf.TSRCell
diff --git a/internal/deepdoc/parser/pdf/table/table_post.go b/internal/deepdoc/parser/pdf/table/table_post.go
index 22eb98bfef..36aff2ed2a 100644
--- a/internal/deepdoc/parser/pdf/table/table_post.go
+++ b/internal/deepdoc/parser/pdf/table/table_post.go
@@ -589,7 +589,15 @@ func mergeFigureGroups(boxes []pdf.TextBox, groups map[figKey][]int, removeSet m
anchor := indices[0]
for _, idx := range indices[1:] {
b := boxes[idx]
- boxes[anchor].Text += "\n" + b.Text
+ if b.Text != "" {
+ if boxes[anchor].Text != "" {
+ boxes[anchor].Text += "\n"
+ }
+ boxes[anchor].Text += b.Text
+ }
+ if boxes[anchor].FigureBounds == nil {
+ boxes[anchor].FigureBounds = b.FigureBounds
+ }
boxes[anchor].X0 = math.Min(boxes[anchor].X0, b.X0)
boxes[anchor].X1 = math.Max(boxes[anchor].X1, b.X1)
boxes[anchor].Top = math.Min(boxes[anchor].Top, b.Top)
@@ -600,14 +608,14 @@ func mergeFigureGroups(boxes []pdf.TextBox, groups map[figKey][]int, removeSet m
}
// ConsolidateFigures merges figure boxes that share the same LayoutNo
-// (i.e., belong to the same DLA figure region) into a single pdf.TextBox.
-// Matches Python's _extract_table_figure + insert_table_figures which pops
-// individual figure boxes and re-inserts one consolidated figure block
-// per DLA region with combined text.
+// into one block, using the matched DLA region for the complete image bounds.
//
// Figure boxes whose text matches the data-source discard pattern
-// (r"(数据|资料|图表)*来源[:: ]") are removed entirely — matching Python's
-// _extract_table_figure discard behavior (pdf_parser.py:1050-1052).
+// (r"(数据|资料|图表)*来源[:: ]") have that text removed, matching Python's
+// _extract_table_figure discard behavior (pdf_parser.py:1050-1052). The
+// region's geometry box survives, so the figure image is preserved even when
+// the attribution line was its only text — a deliberate divergence from
+// Python, which drops the whole region once the box is popped.
func ConsolidateFigures(boxes []pdf.TextBox) []pdf.TextBox {
removeSet := markDataSourceBoxesForRemoval(boxes)
groups := groupFigureBoxes(boxes, removeSet)
@@ -616,9 +624,28 @@ func ConsolidateFigures(boxes []pdf.TextBox) []pdf.TextBox {
mergeFigureGroups(boxes, groups, removeSet)
}
+ applyFigureBounds(boxes)
return FilterBoxesByRemoveSet(boxes, removeSet)
}
+// applyFigureBounds pins every figure box to its matched DLA region.
+//
+// The DLA region is the crop contract: a figure section's geometry must be
+// the region the layout model detected, not the union of the text boxes inside
+// it. Merging unions the member boxes, and NaiveVerticalMerge's empty-box
+// branch can stretch a standalone geometry box (prev.Bottom = curr.Bottom), so
+// the region is re-applied after merging and for single-member groups too.
+func applyFigureBounds(boxes []pdf.TextBox) {
+ for i := range boxes {
+ region := boxes[i].FigureBounds
+ if region == nil {
+ continue
+ }
+ boxes[i].X0, boxes[i].X1 = region.Left, region.Right
+ boxes[i].Top, boxes[i].Bottom = region.Top, region.Bottom
+ }
+}
+
// boxOverlapsPositionPage reports whether a pdf.TextBox overlaps a
// pdf.Position, additionally requiring the box to live on a page the position
// spans. Table positions and table-layout boxes are both stored in page-local
diff --git a/internal/deepdoc/parser/pdf/table/table_post_test.go b/internal/deepdoc/parser/pdf/table/table_post_test.go
index ddf95af8ba..9fc8d74e51 100644
--- a/internal/deepdoc/parser/pdf/table/table_post_test.go
+++ b/internal/deepdoc/parser/pdf/table/table_post_test.go
@@ -749,6 +749,48 @@ func TestConsolidateFigures_OnlyDataSource(t *testing.T) {
}
}
+// TestConsolidateFigures_AppliesFigureBounds pins the crop contract: every
+// surviving figure box is pinned to its DLA region, including a single-member
+// group whose bounds the layout stages stretched (NaiveVerticalMerge's
+// empty-box branch sets prev.Bottom = curr.Bottom) and a merged group whose
+// union differs from the region.
+func TestConsolidateFigures_AppliesFigureBounds(t *testing.T) {
+ boxes := []pdf.TextBox{
+ // Single-member geometry group stretched past its region's bottom.
+ {
+ LayoutType: pdf.LayoutTypeFigure, PageNumber: 0, LayoutNo: "figure-0",
+ X0: 10, X1: 200, Top: 10, Bottom: 55,
+ FigureBounds: &pdf.Position{Left: 10, Right: 200, Top: 10, Bottom: 50},
+ },
+ // Merged group: the caption's union is narrower than the region.
+ {
+ Text: "caption", LayoutType: pdf.LayoutTypeFigure, PageNumber: 0, LayoutNo: "figure-1",
+ X0: 10, X1: 100, Top: 100, Bottom: 120,
+ },
+ {
+ LayoutType: pdf.LayoutTypeFigure, PageNumber: 0, LayoutNo: "figure-1",
+ X0: 10, X1: 200, Top: 100, Bottom: 160,
+ FigureBounds: &pdf.Position{Left: 10, Right: 200, Top: 100, Bottom: 160},
+ },
+ }
+
+ result := ConsolidateFigures(boxes)
+ if len(result) != 2 {
+ t.Fatalf("expected 2 figures, got %d: %+v", len(result), result)
+ }
+ if got := result[0]; got.X0 != 10 || got.X1 != 200 || got.Top != 10 || got.Bottom != 50 {
+ t.Errorf("single-member figure not pinned to its region: got (%v,%v,%v,%v)",
+ got.X0, got.Top, got.X1, got.Bottom)
+ }
+ if got := result[1]; got.X0 != 10 || got.X1 != 200 || got.Top != 100 || got.Bottom != 160 {
+ t.Errorf("merged figure not pinned to its region: got (%v,%v,%v,%v)",
+ got.X0, got.Top, got.X1, got.Bottom)
+ }
+ if result[1].Text != "caption" {
+ t.Errorf("merged figure lost its text: %q", result[1].Text)
+ }
+}
+
func TestExtractTableAndReplace_MergeTablesAcrossPages(t *testing.T) {
// Regression test: ExtractTableAndReplace must correctly replace an
// ALREADY-merged cross-page table (merged upstream by Parser.buildLayout at
diff --git a/internal/deepdoc/parser/pdf/table_extract.go b/internal/deepdoc/parser/pdf/table_extract.go
index c7c2d7e9e6..c3fe0cb014 100644
--- a/internal/deepdoc/parser/pdf/table_extract.go
+++ b/internal/deepdoc/parser/pdf/table_extract.go
@@ -62,7 +62,7 @@ const (
//
// Returns:
// - annotated: page boxes after DLA/TSR annotation write-back (LayoutType,
-// LayoutNo, R/C/H/SP fields) — same length as input pageBoxes.
+// LayoutNo, R/C/H/SP fields), plus image geometry.
// - tables: table candidates detected on this page.
// - dlaRegions: page-local DLA regions payload.
func (p *Parser) enrichOnePageWithDeepDoc(ctx context.Context,
@@ -89,6 +89,10 @@ func (p *Parser) enrichOnePageWithDeepDoc(ctx context.Context,
// caller should use downstream for layout/text-merge.
annotated = append([]pdf.TextBox(nil), pageBoxes...)
annotated = tbl.AnnotateBoxLayouts(annotated, regions, scale, float64(pageImg.Bounds().Dy()))
+ for i := range annotated {
+ annotated[i].PageNumber = pg
+ annotated[i].HasPageNumber = true
+ }
tableMatches := tbl.MatchTableRegions(annotated, regions, scale)
var items []pdf.TableItem
diff --git a/internal/deepdoc/parser/type/types.go b/internal/deepdoc/parser/type/types.go
index cd87566e23..f590d854ec 100644
--- a/internal/deepdoc/parser/type/types.go
+++ b/internal/deepdoc/parser/type/types.go
@@ -117,8 +117,10 @@ type TextBox struct {
Pages []int
LayoutType string
LayoutNo string
- ColID int
- R int
+ // FigureBounds retains DLA geometry in PDF points until figure consolidation.
+ FigureBounds *Position
+ ColID int
+ R int
// IsOCR marks a box produced by an OCR pass (ocrDetectAndRecognize /
// ocrMergeChars), as opposed to one built from embedded PDF chars
// (CharsToBoxes). It scopes OCR-only post-processing (see layout.Dedup*)
diff --git a/internal/parser/parser/pdf_parser_common_test.go b/internal/parser/parser/pdf_parser_common_test.go
index 9dccb609dc..66cf76e611 100644
--- a/internal/parser/parser/pdf_parser_common_test.go
+++ b/internal/parser/parser/pdf_parser_common_test.go
@@ -475,7 +475,7 @@ func TestPDFParseResultToJSON_NoInlineRetainsPositions(t *testing.T) {
PageHeight: map[int]float64{0: 100},
Sections: []deepdoctype.Section{
{
- Text: "Figure caption",
+ Text: "",
LayoutType: deepdoctype.LayoutTypeFigure,
PositionTag: "@@0\t10\t50\t10\t50##",
Positions: []deepdoctype.Position{
@@ -512,12 +512,19 @@ func TestPDFParseResultToJSON_NoInlineRetainsPositions(t *testing.T) {
t.Fatalf("JSON len = %d, want 3", len(res.JSON))
}
- // Figure must retain positions but NOT inline an image.
- figPos, ok := res.JSON[0].ToMap()["_pdf_positions"].([][]any)
+ // A textless figure must remain an image crop item without inlining an image.
+ figure := res.JSON[0].ToMap()
+ if got, want := figure["doc_type_kwd"], "image"; got != want {
+ t.Fatalf("textless figure doc_type_kwd = %v, want %v", got, want)
+ }
+ figPos, ok := figure["_pdf_positions"].([][]any)
if !ok || len(figPos) == 0 {
- t.Fatalf("Figure should retain _pdf_positions, got %v", res.JSON[0].ToMap()["_pdf_positions"])
+ t.Fatalf("textless figure should retain _pdf_positions, got %v", figure["_pdf_positions"])
+ }
+ if !HasPDFCropRegion(figure) {
+ t.Fatal("textless figure should carry a crop region for VLM and chunker")
}
- if img, _ := res.JSON[0].ToMap()["image"].(string); img != "" {
+ if img, _ := figure["image"].(string); img != "" {
t.Fatalf("Figure should not inline image under cgo, got %q", img)
}