fix(docparser): normalize PaddleOCR-VL HTML tables (#1725)

PaddleOCR-VL renders tables as HTML with per-cell text-align styles.
These waste tokens on layout markup and bypass the chunker's
table-protection logic (which only recognizes Markdown pipe tables),
causing large tables to be split mid-row.

Convert inline HTML <table> blocks to GFM Markdown tables when possible,
and strip presentational attributes (style/class/align/...) for tables
using rowspan/colspan that Markdown cannot express.

Refs: https://github.com/Tencent/WeKnora/issues/1725

internal/infrastructure/docparser/html_table_normalizer.go

@@ -0,0 +1,77 @@
+package docparser
+
+import (
+	"regexp"
+	"strings"
+
+	"github.com/JohannesKaufmann/html-to-markdown/v2/converter"
+	"github.com/JohannesKaufmann/html-to-markdown/v2/plugin/base"
+	"github.com/JohannesKaufmann/html-to-markdown/v2/plugin/commonmark"
+	"github.com/JohannesKaufmann/html-to-markdown/v2/plugin/table"
+)
+
+var (
+	// htmlTableBlockPattern matches a single (non-nested) <table>...</table>
+	// block, the form OCR/layout engines such as PaddleOCR-VL emit tables in.
+	htmlTableBlockPattern = regexp.MustCompile(`(?is)<table\b[^>]*>.*?</table>`)
+
+	// htmlLayoutAttrPattern matches presentational HTML attributes that carry
+	// no semantic value (text-align styles, CSS classes, sizing). Structural
+	// attributes like rowspan/colspan are intentionally excluded.
+	htmlLayoutAttrPattern = regexp.MustCompile(
+		`(?is)\s+(?:style|class|align|valign|width|height|bgcolor)\s*=\s*(?:"[^"]*"|'[^']*'|[^\s>]+)`,
+	)
+
+	// htmlSpanAttrPattern detects rowspan/colspan, which Markdown tables cannot
+	// represent; such tables keep their HTML form (attributes stripped) instead.
+	htmlSpanAttrPattern = regexp.MustCompile(`(?i)\b(?:row|col)span\b`)
+
+	// markdownTableSeparatorPattern matches the |---|---| delimiter row that a
+	// valid GFM table must contain.
+	markdownTableSeparatorPattern = regexp.MustCompile(`(?m)^\s*\|?\s*:?-+:?\s*(?:\|\s*:?-+:?\s*)+\|?\s*$`)
+)
+
+// normalizeHTMLTables rewrites inline HTML <table> blocks embedded in OCR
+// markdown output. PaddleOCR-VL emits tables as HTML with per-cell text-align
+// styles, which (1) waste tokens on layout markup and (2) are not recognized
+// by the chunker's table-protection logic, so large tables get split mid-row.
+//
+// Each table block is converted to a GFM Markdown table when possible. Tables
+// that use rowspan/colspan (which Markdown cannot express) fall back to having
+// their presentational attributes stripped so they stay intact as HTML.
+func normalizeHTMLTables(md string) string {
+	if !strings.Contains(strings.ToLower(md), "<table") {
+		return md
+	}
+
+	conv := converter.NewConverter(
+		converter.WithPlugins(
+			base.NewBasePlugin(),
+			commonmark.NewCommonmarkPlugin(),
+			table.NewTablePlugin(),
+		),
+	)
+
+	return htmlTableBlockPattern.ReplaceAllStringFunc(md, func(block string) string {
+		if htmlSpanAttrPattern.MatchString(block) {
+			return stripHTMLLayoutAttrs(block)
+		}
+		converted, err := conv.ConvertString(block)
+		if err != nil {
+			return stripHTMLLayoutAttrs(block)
+		}
+		converted = strings.TrimSpace(converted)
+		if converted == "" || !markdownTableSeparatorPattern.MatchString(converted) {
+			return stripHTMLLayoutAttrs(block)
+		}
+		// Pad with blank lines so the Markdown table is a standalone block that
+		// the chunker recognizes (and protects) as a table.
+		return "\n\n" + converted + "\n\n"
+	})
+}
+
+// stripHTMLLayoutAttrs removes presentational attributes from an HTML fragment
+// while preserving structural attributes (rowspan/colspan) and text content.
+func stripHTMLLayoutAttrs(html string) string {
+	return htmlLayoutAttrPattern.ReplaceAllString(html, "")
+}

internal/infrastructure/docparser/html_table_normalizer_test.go

@@ -0,0 +1,66 @@
+package docparser
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestNormalizeHTMLTables_ConvertsStyledTableToMarkdown(t *testing.T) {
+	// Mirrors PaddleOCR-VL output: a data table where every cell carries a
+	// text-align style that wastes tokens.
+	input := `# 报告
+
+<table><tr><td style="text-align:center;">指标</td><td style="text-align:center;">数值</td></tr>` +
+		`<tr><td style="text-align:center;">营收</td><td style="text-align:right;">10亿</td></tr>` +
+		`<tr><td style="text-align:center;">利润</td><td style="text-align:right;">2.3亿</td></tr></table>
+
+结尾。`
+
+	got := normalizeHTMLTables(input)
+
+	if strings.Contains(got, "<table") {
+		t.Fatalf("expected HTML table to be converted away, got:\n%s", got)
+	}
+	if strings.Contains(got, "text-align") {
+		t.Fatalf("expected style attributes removed, got:\n%s", got)
+	}
+	if !markdownTableSeparatorPattern.MatchString(got) {
+		t.Fatalf("expected a Markdown table separator row, got:\n%s", got)
+	}
+	for _, want := range []string{"指标", "数值", "营收", "10亿", "利润", "2.3亿"} {
+		if !strings.Contains(got, want) {
+			t.Fatalf("expected converted table to retain %q, got:\n%s", want, got)
+		}
+	}
+	if !strings.Contains(got, "# 报告") || !strings.Contains(got, "结尾。") {
+		t.Fatalf("expected surrounding markdown to be preserved, got:\n%s", got)
+	}
+}
+
+func TestNormalizeHTMLTables_StripsAttrsOnSpanTables(t *testing.T) {
+	// rowspan/colspan cannot be expressed in Markdown, so the table stays HTML
+	// but its presentational attributes are stripped.
+	input := `<table><tr><td colspan="2" style="text-align:center;" class="hdr">合计</td></tr>` +
+		`<tr><td style="text-align:left;">A</td><td width="80">B</td></tr></table>`
+
+	got := normalizeHTMLTables(input)
+
+	if !strings.Contains(got, "<table") {
+		t.Fatalf("expected span table to remain HTML, got:\n%s", got)
+	}
+	if !strings.Contains(got, `colspan="2"`) {
+		t.Fatalf("expected colspan to be preserved, got:\n%s", got)
+	}
+	for _, banned := range []string{"text-align", "class=", "width="} {
+		if strings.Contains(got, banned) {
+			t.Fatalf("expected %q to be stripped, got:\n%s", banned, got)
+		}
+	}
+}
+
+func TestNormalizeHTMLTables_NoTableUnchanged(t *testing.T) {
+	input := "# 标题\n\n普通段落，没有表格。\n\n| a | b |\n| --- | --- |\n| 1 | 2 |"
+	if got := normalizeHTMLTables(input); got != input {
+		t.Fatalf("expected content without HTML tables to be unchanged, got:\n%s", got)
+	}
+}

internal/infrastructure/docparser/paddleocr_vl_converter.go

@@ -58,6 +58,12 @@ func (c *PaddleOCRVLReader) Read(ctx context.Context, req *types.ReadRequest) (*
 		return nil, fmt.Errorf("PaddleOCR-VL layout-parsing: %w", err)
 	}
 
+	// PaddleOCR-VL renders tables as styled HTML (per-cell text-align), which
+	// wastes tokens and defeats the chunker's table-protection logic. Convert
+	// them to Markdown tables (or strip layout attributes when conversion is
+	// not possible) before downstream processing.
+	mdContent = normalizeHTMLTables(mdContent)
+
 	imageRefs, mdContent := c.processImages(mdContent, imagesB64)
 	mdContent, imageRefs = ensureOriginalImageRef(req, mdContent, imageRefs)