A fast, helpful, and open-source document parser
Don't end ruled tables at a lone mid-table note row
Logan Markewich0982ef57792a
diff --git a/crates/liteparse/src/markdown_layout/tables.rs b/crates/liteparse/src/markdown_layout/tables.rsindex e00ed29..0c497f6 100644--- a/crates/liteparse/src/markdown_layout/tables.rs+++ b/crates/liteparse/src/markdown_layout/tables.rs@@ -1,3 +1,5 @@+use std::collections::HashSet;+use crate::types::{GraphicPrimitive, ProjectedLine, Rect, TextItem};use super::blocks::{Block, Cell};@@ -4183,6 +4185,77 @@ fn assign_cells(// is a no-op there.let mut line_order: Vec<usize> = (0..lines.len()).collect();line_order.sort_by(|&a, &b| lines[a].bbox.y.total_cmp(&lines[b].bbox.y));++ // Running prose ends the table, but a full-width merged note row inside a+ // ruled table ("(1) Amounts reflect the restatement described in Note 4,+ // ...") has the same shape and is followed by further cell-aligned rows;+ // stopping there would spill the rest of the table into loose lines. So a+ // single prose line only stops the table below the last cell-aligned rows.+ let prose_stop = if stop_at_running_prose {+ let in_grid_row = |idx: usize| {+ let line = &lines[idx];+ let cy = line.bbox.y + line.bbox.height * 0.5;+ if cy < ys[0] || cy > ys[n_rows] || line.text.trim().is_empty() {+ return None;+ }+ find_bucket(ys, cy)+ };+ // A table row: text inside the grid, spread over 2+ columns, with no+ // span straddling an interior boundary. A prose line's runs cut+ // through the boundaries instead.+ let is_cell_aligned_row = |line: &ProjectedLine| {+ let line_x1 = line.bbox.x + line.bbox.width;+ if line.bbox.x < xs[0] - GRID_X_SLACK_PT || line_x1 > xs[n_cols] + GRID_X_SLACK_PT {+ return false;+ }+ let mut cols = HashSet::new();+ for span in line.spans.iter().filter(|s| !s.text.trim().is_empty()) {+ let m0 = (span.x + STRADDLE_MARGIN_PT).clamp(xs[0], xs[n_cols]);+ let m1 = (span.x + span.width - STRADDLE_MARGIN_PT).clamp(xs[0], xs[n_cols]);+ if m1 > m0 && find_bucket(xs, m0) != find_bucket(xs, m1) {+ return false;+ }+ let cx = (span.x + span.width * 0.5).clamp(xs[0], xs[n_cols]);+ cols.extend(find_bucket(xs, cx));+ }+ cols.len() >= 2+ };+ let eligible: Vec<(usize, usize)> = line_order+ .iter()+ .filter_map(|&idx| in_grid_row(idx).map(|row| (idx, row)))+ .collect();+ // Require two table rows below a candidate: a single trailing line+ // in 2+ columns is as likely a page footer as a table row.+ let mut aligned_seen = 0usize;+ let mut first_candidate = eligible.len();+ for (pos, &(idx, _)) in eligible.iter().enumerate().rev() {+ if is_cell_aligned_row(&lines[idx]) {+ aligned_seen += 1;+ if aligned_seen >= 2 {+ break;+ }+ }+ first_candidate = pos;+ }+ // Above that cutoff, still stop when a second prose line follows: a+ // paragraph between two tables sharing one ruled grid, or a form page+ // interleaving prose with checkbox rows. A lone note row has neither.+ let mut prose_positions = eligible+ .iter()+ .enumerate()+ .filter(|&(_, &(idx, row))| {+ should_stop_at_running_prose(&lines[idx], xs, row, idx, dbg)+ })+ .map(|(pos, _)| pos);+ match (prose_positions.next(), prose_positions.next()) {+ (Some(pos), Some(_)) => Some(eligible[pos].0),+ (Some(pos), None) if pos >= first_candidate => Some(eligible[pos].0),+ _ => None,+ }+ } else {+ None+ };+for idx in line_order {let line = &lines[idx];let cy = line.bbox.y + line.bbox.height * 0.5;@@ -4200,7 +4273,7 @@ fn assign_cells(if line.text.trim().is_empty() {continue;}- if stop_at_running_prose && should_stop_at_running_prose(line, xs, row, idx, dbg) {+ if prose_stop == Some(idx) {if dbg {eprintln!("[ruled] stop at running prose line idx={idx} row={row}: {:?}",@@ -5793,6 +5866,89 @@ mod tests {);}+ fn at_y(mut line: ProjectedLine, y: f32) -> ProjectedLine {+ line.bbox.y = y;+ for span in &mut line.spans {+ span.y = y;+ for word in &mut span.words {+ word.y = y;+ }+ }+ line+ }++ fn data_row(cells: &[&str], y: f32) -> ProjectedLine {+ let spans: Vec<_> = cells+ .iter()+ .enumerate()+ .map(|(c, text)| {+ let words = [*text];+ prose_test_span(+ &prose_test_words(&words, 10.0 + 100.0 * c as f32, 5.0, 2.0),+ 10.0,+ )+ })+ .collect();+ at_y(prose_test_line(&cells.join(" "), spans, 10.0), y)+ }++ fn prose_row(y: f32) -> ProjectedLine {+ let text = "During the year ended September 30, 2025, compensation expense increased $1.3 billion, or 2.8%, compared";+ let words: Vec<&str> = text.split_whitespace().collect();+ let positions = prose_test_words(&words, 10.0, 3.0, 2.0);+ at_y(+ prose_test_line(text, vec![prose_test_span(&positions, 10.0)], 10.0),+ y,+ )+ }++ #[test]+ fn merged_note_row_mid_table_keeps_rows_below() {+ let lines = vec![+ data_row(&["Cash", "1,000", "900"], 5.0),+ prose_row(25.0),+ data_row(&["Debt", "500", "600"], 45.0),+ data_row(&["Equity", "500", "300"], 65.0),+ ];+ let xs = [0.0, 100.0, 200.0, 330.0];+ let ys = [0.0, 20.0, 40.0, 60.0, 80.0];++ assert!(should_stop_at_running_prose(&lines[1], &xs, 1, 1, false));+ let (_, consumed) = assign_cells(&lines, &xs, &ys, true, false).unwrap();+ assert_eq!(consumed, [0, 1, 2, 3]);+ }++ #[test]+ fn prose_paragraph_between_tables_stops() {+ let lines = vec![+ data_row(&["Cash", "1,000", "900"], 5.0),+ prose_row(25.0),+ prose_row(45.0),+ data_row(&["Debt", "500", "600"], 65.0),+ data_row(&["Equity", "500", "300"], 85.0),+ ];+ let xs = [0.0, 100.0, 200.0, 330.0];+ let ys = [0.0, 20.0, 40.0, 60.0, 80.0, 100.0];++ let (_, consumed) = assign_cells(&lines, &xs, &ys, true, false).unwrap();+ assert_eq!(consumed, [0]);+ }++ #[test]+ fn running_prose_below_table_still_stops() {+ let lines = vec![+ data_row(&["Cash", "1,000", "900"], 5.0),+ data_row(&["Debt", "500", "600"], 25.0),+ prose_row(45.0),+ prose_row(65.0),+ ];+ let xs = [0.0, 100.0, 200.0, 330.0];+ let ys = [0.0, 20.0, 40.0, 60.0, 80.0];++ let (_, consumed) = assign_cells(&lines, &xs, &ys, true, false).unwrap();+ assert_eq!(consumed, [0, 1]);+ }+#[test]fn long_left_column_label_with_right_column_numbers_continues() {let label = "The subject property benefits from recent renovations and convenient access to neighborhood services.";