Export: auto-bolded metric bullets extract one word per line in stream-order parsers (pypdf/Drive)

クローズ
  • 何か問題がありますか？
#1,193 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

@gaal-agent がすでに取り組んでいます。

2026年10月13日 から。

  • #1196 @gaal-agent による — オープン

評価

難易度
3/5
見積もり時間
1〜2日
初心者へのやさしさ
25/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
python, typescript
領域
frontend

調査の方向性

Start with Layout.drawRunWords in src/lib/pdf/render-ats-pdf.ts (around line 1809), which issues one drawText per word chunk, and check how drawHeaderRuns and wrapRuns feed it. Reproduce by exporting a bullet with auto-bolded metrics and running pypdf's extract_text on it. Done means mixed-weight lines write real spaces, a stream-order extraction returns no single-word lines, and wrapping and page fit are unchanged. Open PR #1196 already targets this, so check it first.

索引モデルが issue の本文から書いたものです。

説明

batch:export-standards bug gaal ready-for-agent ux:edit-export

Problem

Any bullet that contains an auto-bolded metric (autoBoldMetrics, #425) is exported so that stream-order PDF text extractors (pypdf, and Google Drive's text extraction) read it as one word per line:

•
Built
the
org
to
18
engineers
...

Layout-aware extractors (poppler pdftotext) put the line back together correctly, so the export looks fine in most viewers and passes our own round-trip. But ATS parsers in the pypdf family will see a shredded bullet, and those are exactly the bullets that carry the numbers.

Bullets with no metric are unaffected because they take the plain drawText path. Achievement headers that use the emphasis sentinels (Patent ·, Exit ·, …) are affected the same way, through drawHeaderRuns.

Root cause

Layout.drawRunWords (src/lib/pdf/render-ats-pdf.ts, ~L1809) issues one page.drawText per word chunk and advances x by the measured space width. No U+0020 is ever written into the content stream, so the gap between words exists only as positioning. Stream-order extractors treat each separate text-show as its own line.

Repro (synthetic)

  1. Export any résumé with a bullet like Grew the team to 18 engineers and cut p95 latency 6x by rewriting the cache.
  2. python3 -c "import pypdf,sys; print(pypdf.PdfReader(sys.argv[1]).pages[0].extract_text())" out.pdf
  3. That bullet prints one word per line; pdftotext out.pdf - prints it correctly.

On a real 2-page résumé: 197 single-word lines before the fix.

Interim fix (validated locally, not committed)

Merge consecutive same-weight chunks into one drawn string with real spaces, and keep the per-run x from wrapRuns:

const segments: Array<{ str: string; bold: boolean; x: number }> = [];
let x = startX;
for (let w = 0; w < words.length; w++) {
  if (w > 0) { segments[segments.length - 1].str += " "; x += space; }
  for (const chunk of words[w]) {
    const last = segments[segments.length - 1];
    if (last && last.bold === chunk.bold) last.str += chunk.str;
    else segments.push({ str: chunk.str, bold: chunk.bold, x });
    x += chunk.width;
  }
}
for (const seg of segments) this.page.drawText(seg.str, { x: seg.x, y: this.y - size, size, font: seg.bold ? this.fonts.bold : this.fonts.regular, color });

Liberation Sans has the same space advance in Regular and Bold (278/1000 em), so the drawn width still matches what wrapRuns measured and line breaks are unchanged.

Result on the same résumé: 197 → 28 single-word lines. The ones left are bold/regular boundaries (5.5% / to / 7.9%), where words stay whole but each weight switch still starts a new text object.

Acceptance criteria

  • Mixed-weight lines write real spaces between words into the content stream
  • A pypdf-based extraction of an exported bullet with metrics returns the bullet as words on one line. Ideally that includes bold boundaries, e.g. a single BT/ET block with an inline Tf switch, if pdf-lib allows it; otherwise document the residual.
  • Regression test: export a bullet with auto-bolded metrics, extract it with a stream-order extractor, and assert there are no single-word lines
  • No change to measured wrapping (wrapRuns) or page fit
主要言語
TypeScript
スター
11
フォーク
4
30日以内にマージされた PR はありません

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

offlinecv/OfflineCV のほかの issue

offlinecv/OfflineCV の issue をすべて見る

似ている issue

TypeScript の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。