Skip to content

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

Upgrade office_oxide to v0.1.16 and drop temporary caption/table fixes (#21978)

xugangqianga72dd5fc2bfc
diff --git a/go.mod b/go.mod
index 2fdf58b5ca..eba5b875e8 100644
--- a/go.mod
+++ b/go.mod
@@ -71,7 +71,7 @@ require (
github.com/spf13/viper v1.18.2
github.com/ucloud/ucloud-sandbox-sdk-go v0.0.0-20260807065450-08464aef9ed5
github.com/xuri/excelize/v2 v2.11.1-0.20261003002531-6258dcebc4e2
- github.com/yfedoseev/office_oxide/go v0.1.12
+ github.com/yfedoseev/office_oxide/go v0.1.16
github.com/yfedoseev/pdf_oxide/go v0.3.73
github.com/yuin/goldmark v1.7.1
github.com/zeebo/xxh3 v1.0.2
diff --git a/go.sum b/go.sum
index 3dba24fa76..5d81f9380a 100644
--- a/go.sum
+++ b/go.sum
@@ -609,8 +609,8 @@ github.com/xuri/nfp v0.0.2-0.20250530014748-2ddeb826f9a9 h1:+C0TIdyyYmzadGaL/HBL
github.com/xuri/nfp v0.0.2-0.20250530014748-2ddeb826f9a9/go.mod h1:WwHg+CVyzlv/TX9xqBFXEZAuxOPxn2k1GNHwG41IIUQ=
github.com/yargevad/filepathx v1.0.0 h1:SYcT+N3tYGi+NvazubCNlvgIPbzAk7i7y2dwg3I5FYc=
github.com/yargevad/filepathx v1.0.0/go.mod h1:BprfX/gpYNJHJfc35GjRRpVcwWXS89gGulUIU5tK3tA=
-github.com/yfedoseev/office_oxide/go v0.1.12 h1:QLsiB8Aj7LB42D9Vwu1rslRuZg/vSXc4IkVXCDKQv6o=
-github.com/yfedoseev/office_oxide/go v0.1.12/go.mod h1:YLtMlKUkRCp/Q96wsy7D6yoBKDeJnP66UH+c9Bb+E+M=
+github.com/yfedoseev/office_oxide/go v0.1.16 h1:ZZ3/Z5vlwOunUZ5ZY7XItWg3W8Spcw+C2VJIeSdLlJM=
+github.com/yfedoseev/office_oxide/go v0.1.16/go.mod h1:YLtMlKUkRCp/Q96wsy7D6yoBKDeJnP66UH+c9Bb+E+M=
github.com/yfedoseev/pdf_oxide/go v0.3.73 h1:hRad+X4W0ILW3bCQeCvdUKcoy8yMj2TFS6/XtoMk55I=
github.com/yfedoseev/pdf_oxide/go v0.3.73/go.mod h1:QbJ/nLbez0al2EnqEdEPIlGflFprWmiuUM4mo9rNNOI=
github.com/yosida95/uritemplate/v3 v3.0.2 h1:Ed3Oyj9yrmi9087+NczuL5BwkIc4wvTb5zIM+UJPGz4=
diff --git a/internal/parser/parser/docx_caption_cgo_test.go b/internal/parser/parser/docx_caption_cgo_test.go
index 1987dd1e94..b42f083f71 100644
--- a/internal/parser/parser/docx_caption_cgo_test.go
+++ b/internal/parser/parser/docx_caption_cgo_test.go
@@ -134,11 +134,12 @@ func TestDOCXParser_TableCaptionFromHeading_EndToEnd(t *testing.T) {
}
}
-// TestDOCXParser_CaptionStyledParagraphBecomesTableCaption_EndToEnd pins the
-// Go-side temporary stopgap (docx_caption_scan.go): a "Caption"-styled
-// paragraph immediately above a table is attached as the table caption and is
-// NOT emitted as its own text section. (office_oxide itself never surfaces the
-// paragraph style; the Go parser recovers it by re-parsing document.xml.)
+// TestDOCXParser_CaptionStyledParagraphBecomesTableCaption_EndToEnd exercises
+// the real extraction path: office_oxide 0.1.16 surfaces the Word paragraph
+// style name in the IR "style" field (yfedoseev/office_oxide#455), and the Go
+// parser reads it directly. A "Caption"-styled paragraph immediately above a
+// table is attached as the table caption and is NOT emitted as its own text
+// section.
func TestDOCXParser_CaptionStyledParagraphBecomesTableCaption_EndToEnd(t *testing.T) {
res := docxCaptionParse(t, docxCaptionPara("Caption", "设备流转明细")+docxCaptionTable())
td := docxCaptionFindTable(t, res)
diff --git a/internal/parser/parser/docx_caption_scan.go b/internal/parser/parser/docx_caption_scan.go
deleted file mode 100644
index d95d7280e9..0000000000
--- a/internal/parser/parser/docx_caption_scan.go
+++ /dev/null
@@ -1,171 +0,0 @@
-// Package-level note: this file is a TEMPORARY stopgap, not a permanent fix.
-//
-// Problem it works around: office_oxide's IR JSON (the structured output the
-// Go DOCX parser builds sections from) does NOT expose paragraph style names
-// for non-heading paragraphs. It resolves Word heading styles into
-// type "heading" + level, but a paragraph styled with the built-in "Caption"
-// style (or the Chinese "题注") is emitted as a plain {"type":"paragraph"}
-// with no style field. As a result the Go parser cannot tell a caption
-// paragraph from ordinary body text, so a table's visible caption is lost
-// (see the user-reported case "设备流转明细").
-//
-// The correct, robust fix lives upstream: teach office_oxide to emit the
-// paragraph style name so this detection is unnecessary. That is tracked in
-// yfedoseev/office_oxide#455 ("Expose paragraph style name (w:pStyle) in the
-// IR JSON for non-heading paragraphs").
-//
-// Until that lands and office_oxide is upgraded in this repo, we recover the
-// caption the cheap way: re-parse word/document.xml ourselves and pair a
-// caption-styled paragraph with the table that immediately follows it. This
-// MUST be removed once office_oxide emits paragraph styles — do not extend it,
-// do not rely on it from new code. The position-based pairing (document-body
-// table order vs office_oxide IR table order) is approximate and documented at
-// buildDOCXJSONSectionsCore.
-package parser
-
-import (
- "archive/zip"
- "bytes"
- "encoding/xml"
- "fmt"
- "io"
- "strings"
-)
-
-// docxCaptionStyleNames are the Word paragraph style identifiers that mark a
-// paragraph as a table caption for the temporary detection above. The built-in
-// caption style's styleId is "Caption" in every locale; "题注" covers documents
-// where a caption style was created with the localized display name.
-var docxCaptionStyleNames = map[string]bool{
- "caption": true,
- "题注": true,
-}
-
-func isDOCXCaptionStyle(styleID string) bool {
- return docxCaptionStyleNames[strings.ToLower(strings.TrimSpace(styleID))]
-}
-
-// extractDOCXTableCaptions re-parses word/document.xml to recover table
-// captions. It returns one entry per table in document-body reading order: the
-// trimmed text of the paragraph immediately above the table when that
-// paragraph carries a caption style, otherwise "".
-//
-// It is intentionally limited to the common case: paragraphs and tables that
-// are direct children of w:body, in order. Nested tables (e.g. inside a text
-// box) are ignored, matching how office_oxide flattens the typical document.
-// Errors are returned to the caller, which falls back to no captions.
-//
-// TEMPORARY: see the package note at the top of this file.
-func extractDOCXTableCaptions(docx []byte) ([]string, error) {
- zr, err := zip.NewReader(bytes.NewReader(docx), int64(len(docx)))
- if err != nil {
- return nil, fmt.Errorf("docx caption scan: open zip: %w", err)
- }
- var bodyXML []byte
- for _, f := range zr.File {
- if f.Name == "word/document.xml" {
- rc, err := f.Open()
- if err != nil {
- return nil, fmt.Errorf("docx caption scan: open document.xml: %w", err)
- }
- bodyXML, err = io.ReadAll(rc)
- rc.Close()
- if err != nil {
- return nil, fmt.Errorf("docx caption scan: read document.xml: %w", err)
- }
- break
- }
- }
- if bodyXML == nil {
- return nil, fmt.Errorf("docx caption scan: word/document.xml not found")
- }
-
- dec := xml.NewDecoder(bytes.NewReader(bodyXML))
- var captions []string
- var pendingCaption string
-
- var inBody bool
- var bodyDepth int // depth relative to <w:body>; 1 == direct child of body
- var inPara bool
- var paraIsCaption bool
- var inText bool
- var paraText strings.Builder
-
- flushPara := func() {
- if paraIsCaption {
- pendingCaption = strings.TrimSpace(paraText.String())
- } else {
- // Any non-caption paragraph clears a caption that was pending
- // for the next table, so only an *immediately* preceding caption
- // is attached.
- pendingCaption = ""
- }
- inPara = false
- paraIsCaption = false
- inText = false
- paraText.Reset()
- }
-
- for {
- tok, err := dec.Token()
- if err == io.EOF {
- break
- }
- if err != nil {
- return nil, fmt.Errorf("docx caption scan: xml: %w", err)
- }
- switch t := tok.(type) {
- case xml.StartElement:
- local := t.Name.Local
- if !inBody {
- if local == "body" {
- inBody = true
- bodyDepth = 0
- }
- continue
- }
- bodyDepth++
- switch {
- case bodyDepth == 1 && local == "p":
- inPara = true
- paraIsCaption = false
- inText = false
- paraText.Reset()
- case bodyDepth == 1 && local == "tbl":
- captions = append(captions, pendingCaption)
- pendingCaption = ""
- case inPara && local == "pStyle":
- for _, a := range t.Attr {
- if a.Name.Local == "val" {
- paraIsCaption = isDOCXCaptionStyle(a.Value)
- }
- }
- case inPara && local == "t":
- inText = true
- }
- case xml.CharData:
- if inPara && inText {
- paraText.Write(t)
- }
- case xml.EndElement:
- local := t.Name.Local
- if !inBody {
- continue
- }
- switch {
- case bodyDepth == 1 && local == "p" && inPara:
- flushPara()
- case inPara && local == "t":
- inText = false
- case inPara && local == "pStyle":
- // nothing to reset; paraIsCaption stays until paragraph end
- }
- bodyDepth--
- if local == "body" {
- inBody = false
- }
- }
- }
- // A trailing caption paragraph with no following table is discarded.
- return captions, nil
-}
diff --git a/internal/parser/parser/docx_caption_scan_test.go b/internal/parser/parser/docx_caption_scan_test.go
deleted file mode 100644
index 9107069c6c..0000000000
--- a/internal/parser/parser/docx_caption_scan_test.go
+++ /dev/null
@@ -1,141 +0,0 @@
-package parser
-
-import (
- "archive/zip"
- "bytes"
- "testing"
-)
-
-// buildSampleDOCX constructs a minimal but valid .docx (ZIP/OOXML) in memory
-// from the given body children of <w:body>. The caller supplies the inner XML
-// of <w:body> (paragraphs and tables). A default styles.xml is included so the
-// document is well-formed; style names are referenced by w:pStyle w:val.
-func buildSampleDOCX(t *testing.T, bodyInner string) []byte {
- t.Helper()
- document := `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>` +
- `<w:document xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">` +
- `<w:body>` + bodyInner + `</w:body></w:document>`
- styles := `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>` +
- `<w:styles xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">` +
- `<w:style w:type="paragraph" w:default="1" w:styleId="Normal"><w:name w:val="Normal"/></w:style>` +
- `<w:style w:type="paragraph" w:styleId="Heading1"><w:name w:val="heading 1"/></w:style>` +
- `<w:style w:type="paragraph" w:styleId="Caption"><w:name w:val="caption"/></w:style>` +
- `<w:style w:type="paragraph" w:styleId="题注"><w:name w:val="caption"/></w:style>` +
- `</w:styles>`
- contentTypes := `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>` +
- `<Types xmlns="http://schemas.openxmlformats.org/package/2006/content-types">` +
- `<Default Extension="rels" ContentType="application/vnd.openxmlformats-package.relationships+xml"/>` +
- `<Default Extension="xml" ContentType="application/xml"/>` +
- `<Override PartName="/word/document.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.document.main+xml"/>` +
- `<Override PartName="/word/styles.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.styles+xml"/>` +
- `</Types>`
- rels := `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>` +
- `<Relationships xmlns="http://schemas.openxmlformats.org/package/2006/relationships">` +
- `<Relationship Id="rId1" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/officeDocument" Target="word/document.xml"/>` +
- `</Relationships>`
- docRels := `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>` +
- `<Relationships xmlns="http://schemas.openxmlformats.org/package/2006/relationships">` +
- `<Relationship Id="rId1" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/styles" Target="styles.xml"/>` +
- `</Relationships>`
-
- var buf bytes.Buffer
- zw := zip.NewWriter(&buf)
- files := map[string]string{
- "[Content_Types].xml": contentTypes,
- "_rels/.rels": rels,
- "word/document.xml": document,
- "word/styles.xml": styles,
- "word/_rels/document.xml.rels": docRels,
- }
- for name, content := range files {
- w, err := zw.Create(name)
- if err != nil {
- t.Fatalf("zip create %s: %v", name, err)
- }
- if _, err := w.Write([]byte(content)); err != nil {
- t.Fatalf("zip write %s: %v", name, err)
- }
- }
- if err := zw.Close(); err != nil {
- t.Fatalf("zip close: %v", err)
- }
- return buf.Bytes()
-}
-
-func docxScanPara(styleID, text string) string {
- return `<w:p xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">` +
- `<w:pPr><w:pStyle w:val="` + styleID + `"/></w:pPr>` +
- `<w:r><w:t>` + text + `</w:t></w:r></w:p>`
-}
-
-func plainPara(text string) string {
- return `<w:p xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">` +
- `<w:r><w:t>` + text + `</w:t></w:r></w:p>`
-}
-
-func docxScanTable() string {
- return `<w:tbl xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">` +
- `<w:tr><w:tc><w:p><w:r><w:t>A1</w:t></w:r></w:p></w:tc>` +
- `<w:tc><w:p><w:r><w:t>B1</w:t></w:r></w:p></w:tc></w:tr></w:tbl>`
-}
-
-// A Caption-styled paragraph immediately above a table is recovered.
-func TestExtractDOCXTableCaptions_CaptionBeforeTable(t *testing.T) {
- doc := buildSampleDOCX(t,
- docxScanPara("Heading1", "设备台账")+
- docxScanPara("Caption", "设备流转明细")+
- docxScanTable()+
- plainPara("普通正文"))
- caps, err := extractDOCXTableCaptions(doc)
- if err != nil {
- t.Fatalf("extractDOCXTableCaptions: %v", err)
- }
- if len(caps) != 1 {
- t.Fatalf("want 1 table caption, got %d: %v", len(caps), caps)
- }
- if caps[0] != "设备流转明细" {
- t.Errorf("caption = %q, want %q", caps[0], "设备流转明细")
- }
-}
-
-// The Chinese caption style "题注" is detected the same way.
-func TestExtractDOCXTableCaptions_ChineseCaption(t *testing.T) {
- doc := buildSampleDOCX(t, docxScanPara("题注", "设备明细表")+docxScanTable())
- caps, err := extractDOCXTableCaptions(doc)
- if err != nil {
- t.Fatalf("extractDOCXTableCaptions: %v", err)
- }
- if len(caps) != 1 || caps[0] != "设备明细表" {
- t.Fatalf("caption = %v, want [设备明细表]", caps)
- }
-}
-
-// A non-caption paragraph between the caption and the table detaches it.
-func TestExtractDOCXTableCaptions_CaptionNotImmediate(t *testing.T) {
- doc := buildSampleDOCX(t,
- docxScanPara("Caption", "设备流转明细")+
- plainPara("中间说明文字")+
- docxScanTable())
- caps, err := extractDOCXTableCaptions(doc)
- if err != nil {
- t.Fatalf("extractDOCXTableCaptions: %v", err)
- }
- if len(caps) != 1 || caps[0] != "" {
- t.Fatalf("caption = %v, want [\"\"]", caps)
- }
-}
-
-// Multiple tables get the caption of their own preceding caption paragraph.
-func TestExtractDOCXTableCaptions_MultipleTables(t *testing.T) {
- doc := buildSampleDOCX(t,
- docxScanPara("Caption", "第一表")+docxScanTable()+
- plainPara("分隔")+
- docxScanPara("Caption", "第二表")+docxScanTable())
- caps, err := extractDOCXTableCaptions(doc)
- if err != nil {
- t.Fatalf("extractDOCXTableCaptions: %v", err)
- }
- if len(caps) != 2 || caps[0] != "第一表" || caps[1] != "第二表" {
- t.Fatalf("captions = %v, want [第一表 第二表]", caps)
- }
-}
diff --git a/internal/parser/parser/docx_caption_test.go b/internal/parser/parser/docx_caption_test.go
index 74f75f46f4..6206b9a063 100644
--- a/internal/parser/parser/docx_caption_test.go
+++ b/internal/parser/parser/docx_caption_test.go
@@ -152,38 +152,3 @@ func TestBuildDOCXJSONSections_TableCaption_Empty(t *testing.T) {
t.Errorf("empty IR should yield no sections, got %d", len(sections))
}
}
-
-// TEMPORARY Caption-style detection (docx_caption_scan.go): an explicit
-// caption supplied for the table overrides the synthesized heading-location
-// caption.
-func TestBuildDOCXJSONSections_TableCaption_ExplicitPreferredOverLocation(t *testing.T) {
- ir := docxCaptionTableIR(
- `{"type":"heading","level":1,"content":[{"type":"text","text":"设备台账"}]}`,
- docxCaptionTableBody,
- )
- sections := buildDOCXJSONSectionsWithCaptions(ir, nil, []string{"设备流转明细"})
- if got := docxCaptionFindTableCaption(t, sections); got != "设备流转明细" {
- t.Errorf("explicit caption = %q, want %q", got, "设备流转明细")
- }
-}
-
-// An explicit caption is attached even when the table sits under no heading.
-func TestBuildDOCXJSONSections_TableCaption_ExplicitWithNoHeading(t *testing.T) {
- ir := docxCaptionTableIR(docxCaptionTableBody)
- sections := buildDOCXJSONSectionsWithCaptions(ir, nil, []string{"设备流转明细"})
- if got := docxCaptionFindTableCaption(t, sections); got != "设备流转明细" {
- t.Errorf("explicit caption = %q, want %q", got, "设备流转明细")
- }
-}
-
-// A blank entry never overrides a valid heading-location caption.
-func TestBuildDOCXJSONSections_TableCaption_BlankDoesNotOverrideLocation(t *testing.T) {
- ir := docxCaptionTableIR(
- `{"type":"heading","level":1,"content":[{"type":"text","text":"设备台账"}]}`,
- docxCaptionTableBody,
- )
- sections := buildDOCXJSONSectionsWithCaptions(ir, nil, []string{""})
- if got := docxCaptionFindTableCaption(t, sections); got != "Table Location: 设备台账" {
- t.Errorf("caption = %q, want %q", got, "Table Location: 设备台账")
- }
-}
diff --git a/internal/parser/parser/docx_ir.go b/internal/parser/parser/docx_ir.go
index 9bc9bd144d..f7bf8113fd 100644
--- a/internal/parser/parser/docx_ir.go
+++ b/internal/parser/parser/docx_ir.go
@@ -357,41 +357,32 @@ func docxHeadingLocationCaption(headings []docxHeading) string {
// heading styles into element type "heading" plus a level, so the outline
// is tracked from those elements alone.
//
-// This is the caption-only (heading hierarchy) entry point. Use
-// buildDOCXJSONSectionsWithCaptions to also attach Caption-styled paragraph
-// text (a temporary stopgap; see docx_caption_scan.go).
+// buildDOCXJSONSections builds sections from an office_oxide IR. The heading
+// hierarchy supplies each table's caption; a "Caption"-styled paragraph that
+// immediately precedes a table overrides that with its own text. (Paragraph-
+// style recovery used to be a temporary stopgap that re-parsed
+// word/document.xml; office_oxide now emits the style directly — see
+// yfedoseev/office_oxide#455.)
//
// A nil sink means a text-only parse: embedded images are skipped and no
// media cache is required. Normal DOCX/DOC parsing always passes a sink.
func buildDOCXJSONSections(irJSON string, sink *officeMediaSink) []map[string]any {
- return buildDOCXJSONSectionsCore(irJSON, sink, nil)
+ return buildDOCXJSONSectionsCore(irJSON, sink)
}
-// buildDOCXJSONSectionsWithCaptions is buildDOCXJSONSections plus the
-// TEMPORARY Caption-styled-paragraph detection. tableCaptions must be aligned
-// by position with the tables in the IR (see docx_caption_scan.go).
-//
-// TEMPORARY: remove once office_oxide emits paragraph styles
-// (yfedoseev/office_oxide#455).
-func buildDOCXJSONSectionsWithCaptions(irJSON string, sink *officeMediaSink, tableCaptions []string) []map[string]any {
- return buildDOCXJSONSectionsCore(irJSON, sink, tableCaptions)
-}
-
-// buildDOCXJSONSectionsCore is the shared implementation of the two entry
-// points above. tableCaptions, when non-nil, supplies a descriptive caption
-// for each table indexed by reading order; an explicit caption is preferred
-// over the synthesized heading-location caption. The position alignment
-// between tableCaptions and the IR tables is approximate (document-body order
-// vs office_oxide IR order) and is a known limitation of the temporary
-// Caption-style stopgap in docx_caption_scan.go.
-func buildDOCXJSONSectionsCore(irJSON string, sink *officeMediaSink, tableCaptions []string) []map[string]any {
+// buildDOCXJSONSectionsCore builds sections from an office_oxide IR. The
+// heading hierarchy supplies each table's caption; a "Caption"-styled
+// paragraph that immediately precedes a table overrides that with its own
+// text. (Paragraph-style recovery used to be a temporary stopgap that
+// re-parsed word/document.xml; office_oxide now emits the style directly.)
+func buildDOCXJSONSectionsCore(irJSON string, sink *officeMediaSink) []map[string]any {
var ir docxIRDocument
if err := json.Unmarshal([]byte(irJSON), &ir); err != nil {
return nil
}
var sections []map[string]any
tableSequence := 0
- captionSequence := 0
+ var pendingCaption string
var headings []docxHeading
var emitElements func([]docxIRElement, bool)
emitElements = func(elements []docxIRElement, topLevel bool) {
@@ -405,57 +396,6 @@ func buildDOCXJSONSectionsCore(irJSON string, sink *officeMediaSink, tableCaptio
return
}
el := elements[ei]
- // TEMPORARY: this merge is a stopgap. Revisit it once docxIRRow
- // carries a three-state header field and office_oxide populates
- // `TableRow.is_header` for .doc tables (upstream issue
- // yfedoseev/office_oxide#456) — at that point prefer rejoining via
- // the real header flag and drop the structural proxy below.
- //
- // office_oxide splits a .doc table whose first row is a styled
- // header/title row into a 1-row "header" table immediately
- // followed by the body table (the legacy binary .doc format
- // stores such a table as two separate Table objects with a blank
- // paragraph between them; walk_paragraphs flushes a new Element::Table
- // at every non-fInTable paragraph, and emit_prose drops the blank
- // one, so the two tables end up adjacent in the IR). Rejoin a
- // 1-row table fragment with the adjacent following table when their
- // first rows carry the same column count, so one logical table is
- // emitted as a single item instead of a header chunk + body chunk
- // (each would otherwise carry the same caption).
- //
- // Why a structural proxy instead of the row's `is_header` flag:
- // office_oxide's IR DOES carry `TableRow.is_header`, but the .doc
- // converter (convert_doc.rs) never populates it — every .doc row
- // arrives with is_header=false, verified on the real document that
- // triggered this bug. Only the .docx converter sets it, and .docx
- // never splits. So `is_header` is a dead signal for .doc and cannot
- // be used to confirm the leading fragment is a header or that the
- // following table lacks one. The 1-row gate is the strongest
- // available proxy: a 1-row table adjacent to a same-width table is
- // overwhelmingly a header fragment, because a genuinely standalone
- // 1-row table immediately followed by another table (with only a
- // blank paragraph between) is rare.
- //
- // Regression guards baked into the merge condition:
- // - leading fragment is exactly 1 row: two genuinely separate
- // MULTI-row tables (same width, blank paragraph between) are
- // never merged.
- // - same first-row column count: differently-shaped tables stay
- // separate.
- // - adjacency (no element between): any heading/paragraph/image
- // in between keeps the two tables distinct.
- if topLevel && el.Type == "table" && len(el.Rows) == 1 && ei+1 < len(elements) {
- if nxt := elements[ei+1]; nxt.Type == "table" &&
- docxIRTableCols(el) > 0 && docxIRTableCols(el) == docxIRTableCols(nxt) {
- merged := nxt
- rows := make([]docxIRRow, 0, len(nxt.Rows)+len(el.Rows))
- rows = append(rows, el.Rows...)
- rows = append(rows, nxt.Rows...)
- merged.Rows = rows
- el = merged
- ei++ // fold the following table into this one; skip it
- }
- }
switch el.Type {
case "paragraph", "heading":
if el.Type == "heading" {
@@ -465,17 +405,13 @@ func buildDOCXJSONSectionsCore(irJSON string, sink *officeMediaSink, tableCaptio
}
headings = pushDOCXHeading(headings, level, docxElementText(el, "\n"))
}
- // TEMPORARY (docx_caption_scan.go): when a paragraph's text
- // exactly matches the Caption recovered for the next table,
- // suppress it as a standalone text section — it is attached
- // to that table as the caption instead. The text match keeps
- // this safe if the document scan and the IR element order
- // ever desync: a non-matching paragraph is never dropped.
- if topLevel && captionSequence < len(tableCaptions) {
- if explicit := strings.TrimSpace(tableCaptions[captionSequence]); explicit != "" &&
- strings.TrimSpace(docxElementText(el, " ")) == explicit {
- continue
- }
+ // A "Caption"-styled paragraph (office_oxide 0.1.16 exposes the
+ // Word style name via the IR "style" field, yfedoseev/office_oxide#455)
+ // is attached to the next table as its caption and is not emitted as a
+ // standalone text section. It is consumed by the next top-level table.
+ if topLevel && strings.EqualFold(strings.TrimSpace(el.Style), "Caption") {
+ pendingCaption = strings.TrimSpace(docxElementText(el, " "))
+ continue
}
sections = appendDOCXParagraphSections(sections, el, el.Type == "heading", sink)
@@ -484,20 +420,15 @@ func buildDOCXJSONSectionsCore(irJSON string, sink *officeMediaSink, tableCaptio
case "table":
tableSequence++
- if topLevel {
- captionSequence++
- }
if len(el.Rows) == 0 {
continue
}
- // Prefer an explicit Caption-styled paragraph (temporary
- // stopgap, see docx_caption_scan.go) over the synthesized
- // heading-location caption.
+ // A "Caption"-styled paragraph that preceded this table (office_oxide#455)
+ // overrides the synthesized heading-location caption.
caption := docxHeadingLocationCaption(headings)
- if idx := captionSequence - 1; topLevel && idx >= 0 && idx < len(tableCaptions) {
- if explicit := strings.TrimSpace(tableCaptions[idx]); explicit != "" {
- caption = explicit
- }
+ if topLevel && pendingCaption != "" {
+ caption = pendingCaption
+ pendingCaption = ""
}
sections = appendOfficeTableSections(sections, el, caption, fmt.Sprintf("docx-table-%d", tableSequence), sink)
@@ -516,6 +447,17 @@ func buildDOCXJSONSectionsCore(irJSON string, sink *officeMediaSink, tableCaptio
for _, sec := range ir.Sections {
emitElements(sec.Elements, true)
}
+
+ // An orphan Caption-styled paragraph (no following table) is emitted as a
+ // normal text section so it is never silently dropped.
+ if pendingCaption != "" {
+ sections = append(sections, map[string]any{
+ "text": pendingCaption,
+ "image": nil,
+ "doc_type_kwd": "text",
+ })
+ }
+
return sections
}
diff --git a/internal/parser/parser/docx_ir_caption_style_test.go b/internal/parser/parser/docx_ir_caption_style_test.go
new file mode 100644
index 0000000000..9f3a783f91
--- /dev/null
+++ b/internal/parser/parser/docx_ir_caption_style_test.go
@@ -0,0 +1,140 @@
+package parser
+
+import (
+ "testing"
+
+ "ragflow/internal/entity"
+)
+
+// docxCountTables returns the number of top-level table sections.
+func docxCountTables(sections []map[string]any) int {
+ n := 0
+ for _, s := range sections {
+ if dt, _ := s["doc_type_kwd"].(string); dt == "table" {
+ n++
+ }
+ }
+ return n
+}
+
+// docxFirstTableCaption returns the Caption of the first table section.
+func docxFirstTableCaption(t *testing.T, sections []map[string]any) string {
+ t.Helper()
+ for _, s := range sections {
+ if dt, _ := s["doc_type_kwd"].(string); dt == "table" {
+ td, ok := s["table"].(*entity.TableData)
+ if !ok || td == nil {
+ t.Fatalf("table section missing *entity.TableData")
+ }
+ return td.Caption
+ }
+ }
+ t.Fatal("no table section found")
+ return ""
+}
+
+// docxHasText reports whether a standalone text section carries exactly want.
+func docxHasText(sections []map[string]any, want string) bool {
+ for _, s := range sections {
+ if dt, _ := s["doc_type_kwd"].(string); dt == "text" {
+ if txt, _ := s["text"].(string); txt == want {
+ return true
+ }
+ }
+ }
+ return false
+}
+
+// A Caption-styled paragraph (office_oxide 0.1.16 emits "style":"Caption")
+// immediately above a table must be attached as the table caption and must
+// NOT be emitted as its own text section. Before office_oxide exposed the
+// style field this required re-parsing word/document.xml (docx_caption_scan.go);
+// now the IR carries it directly.
+func TestBuildDOCXJSONSections_TableCaption_FromStyle(t *testing.T) {
+ ir := `{"sections":[{"elements":[
+ {"type":"paragraph","style":"Caption","content":[{"type":"text","text":"设备流转明细"}]},
+ {"type":"table","rows":[{"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"A1"}]}]},{"content":[{"type":"paragraph","content":[{"type":"text","text":"B1"}]}]}]}]}
+ ]}]}`
+ sections := buildDOCXJSONSections(ir, newOfficeMediaSink(t.Context(), officeMediaFormatDOCX))
+ if got := docxFirstTableCaption(t, sections); got != "设备流转明细" {
+ t.Errorf("caption = %q, want %q", got, "设备流转明细")
+ }
+ if docxHasText(sections, "设备流转明细") {
+ t.Error("Caption-styled paragraph must be consumed as caption, not left as a text section")
+ }
+}
+
+// An explicit Caption-styled paragraph wins over the synthesized
+// heading-location caption when both precede the table.
+func TestBuildDOCXJSONSections_TableCaption_StylePreferredOverHeading(t *testing.T) {
+ ir := `{"sections":[{"elements":[
+ {"type":"heading","level":1,"content":[{"type":"text","text":"设备台账"}]},
+ {"type":"paragraph","style":"Caption","content":[{"type":"text","text":"设备流转明细"}]},
+ {"type":"table","rows":[{"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"A1"}]}]},{"content":[{"type":"paragraph","content":[{"type":"text","text":"B1"}]}]}]}]}
+ ]}]}`
+ sections := buildDOCXJSONSections(ir, newOfficeMediaSink(t.Context(), officeMediaFormatDOCX))
+ if got := docxFirstTableCaption(t, sections); got != "设备流转明细" {
+ t.Errorf("caption = %q, want explicit caption %q", got, "设备流转明细")
+ }
+}
+
+// office_oxide reports built-in Caption as the capitalized name; the match
+// must be case-insensitive so "caption" (or any casing) is still recognized.
+func TestBuildDOCXJSONSections_TableCaption_StyleCaseInsensitive(t *testing.T) {
+ ir := `{"sections":[{"elements":[
+ {"type":"paragraph","style":"caption","content":[{"type":"text","text":"设备流转明细"}]},
+ {"type":"table","rows":[{"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"A1"}]}]}]}]}
+ ]}]}`
+ sections := buildDOCXJSONSections(ir, newOfficeMediaSink(t.Context(), officeMediaFormatDOCX))
+ if got := docxFirstTableCaption(t, sections); got != "设备流转明细" {
+ t.Errorf("caption = %q, want %q", got, "设备流转明细")
+ }
+}
+
+// A Caption-styled paragraph with NO following table is not a caption; it
+// must be emitted as an ordinary text section (never silently dropped).
+func TestBuildDOCXJSONSections_TableCaption_StyleWithoutTableEmittedAsText(t *testing.T) {
+ ir := `{"sections":[{"elements":[
+ {"type":"paragraph","style":"Caption","content":[{"type":"text","text":"孤立题注"}]}
+ ]}]}`
+ sections := buildDOCXJSONSections(ir, newOfficeMediaSink(t.Context(), officeMediaFormatDOCX))
+ if docxCountTables(sections) != 0 {
+ t.Errorf("expected no table sections, got %d", docxCountTables(sections))
+ }
+ if !docxHasText(sections, "孤立题注") {
+ t.Error("orphan Caption-styled paragraph must be emitted as a text section")
+ }
+}
+
+// A plain (non-Caption) paragraph above a table must never become a caption;
+// only the heading hierarchy is used for it.
+func TestBuildDOCXJSONSections_TableCaption_NonCaptionParagraphIgnored(t *testing.T) {
+ ir := `{"sections":[{"elements":[
+ {"type":"paragraph","content":[{"type":"text","text":"设备流转明细"}]},
+ {"type":"table","rows":[{"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"A1"}]}]}]}]}
+ ]}]}`
+ sections := buildDOCXJSONSections(ir, newOfficeMediaSink(t.Context(), officeMediaFormatDOCX))
+ if got := docxFirstTableCaption(t, sections); got == "设备流转明细" {
+ t.Errorf("plain paragraph must not become a caption, got %q", got)
+ }
+}
+
+// Regression for the old .doc split workaround (office_oxide#456): a 1-row
+// table immediately followed by a same-width multi-row table must NOT be
+// merged into one section. office_oxide 0.1.16 no longer splits .doc tables,
+// so the structural proxy in buildDOCXJSONSectionsCore is gone and two
+// adjacent tables stay distinct.
+func TestBuildDOCXJSONSections_AdjacentTablesNotMerged(t *testing.T) {
+ ir := `{"sections":[{"elements":[
+ {"type":"table","rows":[{"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"H"}]}]}]}]},
+ {"type":"table","rows":[
+ {"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"A1"}]}]}]},
+ {"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"A2"}]}]}]},
+ {"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"A3"}]}]}]}
+ ]}
+ ]}]}`
+ sections := buildDOCXJSONSections(ir, newOfficeMediaSink(t.Context(), officeMediaFormatDOCX))
+ if got := docxCountTables(sections); got != 2 {
+ t.Errorf("adjacent tables must stay distinct, got %d tables (want 2)", got)
+ }
+}
diff --git a/internal/parser/parser/docx_ir_test.go b/internal/parser/parser/docx_ir_test.go
index 6ef50aaea8..1d1576b0a6 100644
--- a/internal/parser/parser/docx_ir_test.go
+++ b/internal/parser/parser/docx_ir_test.go
@@ -54,44 +54,6 @@ func TestBuildDOCXJSONSections_EmptyTableSkipped(t *testing.T) {
}
}
-// TestBuildDOCXJSONSections_MergesSplitDocTable pins the fix for the
-// office_oxide .doc artifact where a table whose first row is a styled
-// header is emitted as a 1-row "header" table immediately followed by the
-// body table. The two fragments share the same column count and sit under
-// the same heading, so they must be rejoined into a single table item
-// (otherwise the chunker emits a header chunk + body chunk, each carrying
-// the same caption).
-func TestBuildDOCXJSONSections_MergesSplitDocTable(t *testing.T) {
- const cell = `{"content":[{"type":"paragraph","content":[{"type":"text","text":"x"}]}]}`
- // header: 1 row x 3 cols; body: 2 rows x 3 cols — same width.
- irJSON := fmt.Sprintf(`{"sections":[{"title":"","elements":[
- {"type":"heading","level":1,"content":[{"type":"text","text":"H"}]},
- {"type":"table","rows":[{"cells":[%s,%s,%s]}]},
- {"type":"table","rows":[{"cells":[%s,%s,%s]},{"cells":[%s,%s,%s]}]}
- ]}]}`, cell, cell, cell, cell, cell, cell, cell, cell, cell)
-
- got := buildDOCXJSONSections(irJSON, nil)
- var tables []map[string]any
- for _, it := range got {
- if dt, _ := it["doc_type_kwd"].(string); dt == "table" {
- tables = append(tables, it)
- }
- }
- if len(tables) != 1 {
- t.Fatalf("expected 1 merged table item, got %d: %#v", len(tables), got)
- }
- td, ok := tables[0]["table"].(*entity.TableData)
- if !ok || td == nil {
- t.Fatalf("merged table has no TableData: %#v", tables[0])
- }
- if len(td.Grid) != 3 {
- t.Errorf("merged table should have 3 rows (1 header + 2 body), got %d: %#v", len(td.Grid), td.Grid)
- }
- if td.Caption != "Table Location: H" {
- t.Errorf("merged table caption = %q, want %q", td.Caption, "Table Location: H")
- }
-}
-
// TestBuildDOCXJSONSections_DifferentColumnTablesStaySeparate pins that
// two adjacent tables of different column counts are NOT merged: they are
// genuinely separate tables, not fragments of one.
diff --git a/internal/parser/parser/docx_parser.go b/internal/parser/parser/docx_parser.go
index 780f1d7371..0267dfed21 100644
--- a/internal/parser/parser/docx_parser.go
+++ b/internal/parser/parser/docx_parser.go
@@ -86,14 +86,13 @@ func (p *DOCXParser) ParseWithResult(ctx context.Context, filename string, data
if irJSON == "" {
return ParseResult{Err: fmt.Errorf("docx extract: empty IR, cannot build JSON output")}
}
- // TEMPORARY Caption-styled-paragraph detection (office_oxide does
- // not yet expose paragraph styles; yfedoseev/office_oxide#455).
- // Remove once office_oxide is upgraded to emit them.
- tableCaptions, capErr := extractDOCXTableCaptions(data)
- if capErr != nil {
- tableCaptions = nil
- }
- sections := buildDOCXJSONSectionsWithCaptions(irJSON, sink, tableCaptions)
+ // office_oxide 0.1.16 exposes the Word paragraph style name (e.g.
+ // "Caption") in the IR "style" field (yfedoseev/office_oxide#455); a
+ // Caption-styled paragraph is attached to the table that follows it and is
+ // not emitted as a standalone text section. This replaces the temporary
+ // word/document.xml re-parse that the Go parser used before the field
+ // existed.
+ sections := buildDOCXJSONSections(irJSON, sink)
if sink != nil && sink.err != nil {
return ParseResult{Err: sink.err}
}
diff --git a/internal/parser/parser/table_regression_test.go b/internal/parser/parser/table_regression_test.go
index 216d2435b1..ebbe76e2e4 100644
--- a/internal/parser/parser/table_regression_test.go
+++ b/internal/parser/parser/table_regression_test.go
@@ -26,9 +26,9 @@ func TestTableRegressionMarkdownMixedBackticksKeepTrailingCells(t *testing.T) {
}
func TestTableRegressionDOCXCaptionAfterNestedTable(t *testing.T) {
- ir := `{"sections":[{"elements":[{"type":"text_box","content":[{"type":"table","rows":[{"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"Nested"}]}]}]}]}]},{"type":"paragraph","content":[{"type":"text","text":"Outer Caption"}]},{"type":"table","rows":[{"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"Outer"}]}]}]}]}]}]}`
+ ir := `{"sections":[{"elements":[{"type":"text_box","content":[{"type":"table","rows":[{"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"Nested"}]}]}]}]}]},{"type":"paragraph","style":"Caption","content":[{"type":"text","text":"Outer Caption"}]},{"type":"table","rows":[{"cells":[{"content":[{"type":"paragraph","content":[{"type":"text","text":"Outer"}]}]}]}]}]}]}`
- after := buildDOCXJSONSectionsWithCaptions(ir, nil, []string{"Outer Caption"})
+ after := buildDOCXJSONSections(ir, nil)
if len(after) != 2 {
t.Fatalf("expected two tables without duplicate caption text: %#v", after)
}
@@ -37,7 +37,7 @@ func TestTableRegressionDOCXCaptionAfterNestedTable(t *testing.T) {
if tableRowsContain(td, "Nested") && td.Caption != "" {
t.Error("nested table stole the outer caption")
}
- if tableRowsContain(td, "Outer") && td.Caption != "Outer Caption" {
+ if tableRowsContain(td, "Outer") && td.Caption != "Outer Caption" { // caption from preceding Caption-styled paragraph
t.Error("outer table lost its caption")
}
}