tables: keep wrapped labels in TSR output

napi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@firecrawl/pdf-inspector",
-  "version": "1.8.2",
+  "version": "1.8.3",
   "description": "Fast PDF classification and text extraction. Detect text-based vs scanned PDFs, extract text by region with quality checks. Native Rust performance via napi-rs.",
   "main": "index.js",
   "types": "index.d.ts",

src/lib.rs

@@ -2152,13 +2152,15 @@ fn try_expand_multi_row_cells(
 /// * `phantom_empty_row` — a row whose every cell is empty, surrounded
 ///   above and below by rows with content. SLANet sometimes emits an
 ///   extra row that doesn't correspond to any visible PDF row.
-/// * `multi_row_in_cell` — at least one `rowspan==1` cell encloses
-///   PDF text items that cluster into two distinct visual lines
+/// * `multi_row_in_cell` — at least one non-label `rowspan==1` cell
+///   encloses PDF text items that cluster into two distinct visual lines
 ///   separated by a whitespace gap larger than the line height. Cells
 ///   declared as `rowspan>1` are excluded since they are *expected*
-///   to span multiple lines. SLANet's row under-detection on
-///   tightly-packed tables produces the rowspan==1-but-multi-line
-///   pattern (the FNBO failure mode).
+///   to span multiple lines. First-row/first-column wraps are ignored
+///   unless the in-place row expansion has enough support to repair them,
+///   because those are often legitimate wrapped headers or row labels.
+///   SLANet's row under-detection on tightly-packed tables produces the
+///   rowspan==1-but-multi-line pattern (the FNBO failure mode).
 fn detect_tsr_quality_issue(
     buffer: &[u8],
     input: &TsrTableInput,
@@ -2216,6 +2218,8 @@ fn detect_tsr_quality_issue(
     };
     let expanded_cells =
         try_expand_multi_row_cells(cells, &items, page_h, coords, adaptive_threshold);
+    let first_row = cells.iter().map(|cell| cell.row).min().unwrap_or(0);
+    let first_col = cells.iter().map(|cell| cell.col).min().unwrap_or(0);
 
     for cell in cells {
         // rowspan>1 cells are intentionally multi-line — skip them.
@@ -2229,14 +2233,30 @@ fn detect_tsr_quality_issue(
         if cell_items.len() < 2 {
             continue;
         }
-        if cluster_tsr_cell_text_lines(cell_items).len() >= 2 {
+        if cluster_tsr_cell_text_lines(cell_items).len() < 2 {
+            continue;
+        }
+        if expanded_cells.is_some() {
             return Ok(Some(TsrQualityIssue::MultiRowInCell { expanded_cells }));
         }
+        if !is_wrapped_tsr_label_cell(cell, first_row, first_col) {
+            return Ok(Some(TsrQualityIssue::MultiRowInCell {
+                expanded_cells: None,
+            }));
+        }
     }
 
     Ok(None)
 }
 
+fn is_wrapped_tsr_label_cell(
+    cell: &tables::StructuredCell,
+    first_row: usize,
+    first_col: usize,
+) -> bool {
+    cell.is_header || cell.row == first_row || cell.col == first_col
+}
+
 /// Auto-fallback variant of [`extract_tables_with_structure_mem`]:
 /// runs the TSR-hybrid path, checks the resulting cells for known
 /// SLANet detection pathologies (phantom rows, multi-row-in-cell text),

tests/integration_tests.rs

@@ -2532,6 +2532,61 @@ fn test_auto_expands_under_counted_vector_grid_rows() {
     );
 }
 
+#[test]
+fn test_auto_keeps_wrapped_header_vector_grid_doc51() {
+    use pdf_inspector::{extract_tables_with_structure_auto_mem, TsrTableInput};
+
+    let buf = std::fs::read("tests/fixtures/government_positions_women.pdf").unwrap();
+    let crop = [0.0, 0.0, 612.0, 792.0];
+    let grid = detect_vector_grid_in_region_mem(&buf, 0, crop, 200.0)
+        .unwrap()
+        .expect("expected doc 51 vector grid");
+    assert_eq!(
+        grid.cell_bboxes.len(),
+        36,
+        "doc 51 should have a 9x4 vector grid"
+    );
+
+    let results = extract_tables_with_structure_auto_mem(
+        &buf,
+        &[TsrTableInput {
+            page: 0,
+            crop_pdf_pt_bbox: crop,
+            render_dpi: 200.0,
+            structure_tokens: grid.structure_tokens,
+            cell_bboxes: grid.cell_bboxes,
+        }],
+    )
+    .unwrap();
+
+    assert_eq!(results.len(), 1);
+    let r = &results[0];
+    assert!(
+        r.fallback_reason.is_none(),
+        "wrapped header/label text should not trigger heuristic fallback: {:?}\n{}",
+        r.fallback_reason,
+        r.markdown
+    );
+    let md = &r.markdown;
+    assert!(md.contains("Government Position"), "missing header: {md}");
+    assert!(
+        md.contains("Aquino Administration"),
+        "missing Aquino header: {md}"
+    );
+    assert!(
+        md.contains("Ramos Administration"),
+        "missing Ramos header: {md}"
+    );
+    assert!(
+        md.contains("City Municipal Councilor"),
+        "row label was truncated: {md}"
+    );
+    assert!(
+        !md.contains("|Position||Administration"),
+        "heuristic fallback split the header row: {md}"
+    );
+}
+
 #[test]
 fn test_auto_returns_empty_inputs() {
     use pdf_inspector::extract_tables_with_structure_auto_mem;