pdfcopy

Drag across a wrap. Paste one sentence.

划过自动折行的那几个字，贴出来是一句完整的话。

A PDF is a stack of “draw these outlines at (x, y)” commands. It does not store sentences. Most readers put a \n on the clipboard at every visual wrap, drag in Page 3 of 12, and — when ToUnicode is missing — paste ????.

pdfcopy rebuilds the sentence a person meant to copy, then writes it back as a copy mesh: each visible letter gets an invisible twin at the same box, soft wraps become spaces, and each sentence carries /ActualText. Offline. No GPU.

Try it

No install — the demo desk is live:

https://ch1rpy2613--pdfcopy.modal.run

Or run it yourself in — pdfcopy serve --open in the Codespace terminal (the port forwards automatically).

Install the CLI (Python 3.10+):

pipx install pdfcopy        # or: pip install pdfcopy
pdfcopy extract contract.pdf

Self-host the demo desk:

docker build -t pdfcopy .
docker run --rm -p 8000:8000 pdfcopy

Or one command on Modal (free tier covers it): modal deploy modal_app.py.

Open the demo desk: drag a PDF onto the lit page, or hit the narrow-column sample. Drag across a wrap — amber highlight follows the ink, the note card on the right is the locked sentence. Copy writes text/plain and text/html. Flip to 断行 for the staircase.

What it does

• Infers spaces from glyph gaps (many PDFs never write U+0020)
• Splits reconstructed paragraphs into sentences (CJK 。！？；, Latin .?! with abbrev/decimal guards)
• Joins soft wraps; keeps list items, tables, and code
• Stitches hyphenated line ends (inter- + national)
• Drops running heads and 第 3 页 共 12 页
• Leaves CJK characters unspaced; keeps a space around Latin
• Writes a glyph-aligned copy mesh into a standard PDF so other readers’ Ctrl+C can lock a sentence
• Keeps original ink (Artifact-wrapped); --outline is optional for readers that ignore Artifact
• Also repairs ToUnicode, a structure tree, and a public overlay convention

It does not invent a new format. Acrobat and Preview still open the file. After pdfcopy fix, a two-word drag through a wrap pastes the sentence, not five line fragments.

Commands

pdfcopy extract paper.pdf           # reconstructed text on stdout
pdfcopy extract paper.pdf --naive   # what most readers copy
pdfcopy extract paper.pdf --page 2
pdfcopy fix paper.pdf -o paper.fixed.pdf
pdfcopy fix paper.pdf -o paper.fixed.pdf --outline   # path-out the original letters
pdfcopy info scan.pdf

Scanned pages have no text layer. v1 reports that and stops. It does not OCR.

Corpus

corpus/ holds 20 fixtures — invoice, two-column article, contract, Chinese without ToUnicode, scan, fake spaces — each with expected paste text.

Fixture 13-narrow-column is the proof: typical copy is a staircase plus Page 1 of 1; a bbox over the middle sentence pastes that sentence only, with no \n.

python scripts/generate_corpus.py
pytest
python scripts/make_gif.py

中文说明

划过自动折行的那几个字，贴出来是一句完整的话。

在线演示（无需安装）：https://ch1rpy2613--pdfcopy.modal.run

PDF 里存的不是句子，而是一串“在坐标 (x, y) 画出这些轮廓”的指令。所以普通阅读器复制出来的是按视觉换行切碎的碎片，还夹带“第 3 页 共 12 页”；缺 ToUnicode 的中文 PDF 甚至只贴得出 ????。

pdfcopy 从字形坐标重建出你真正想复制的那句话，再以**复制网格（copy mesh）**写回标准 PDF：每个可见字母在同位置获得一个不可见孪生，软换行变成真正的空格，每句话带 /ActualText。全程离线、无需 GPU，Acrobat / 预览 / 浏览器照常打开。

它做了什么：

• 从字形间距推断空格（很多 PDF 根本不写 U+0020）
• 把重建段落切分成句（中文 。！？；，英文 .?!，带缩写 / 小数保护）
• 合并软换行；保留列表、表格、代码的原始换行
• 拼接行尾连字符（inter- + national）
• 去掉页眉页脚和“第 3 页 共 12 页”
• 中文之间不加空格；中西文之间保留空格
• 修复 ToUnicode、补结构树，重建结果按公开约定写入 overlay

pipx install pdfcopy
pdfcopy extract 合同.pdf
pdfcopy serve --open   # 打开本地演示台

扫描件没有文本层，v1 只报告、不做 OCR。

Status

Sentence lock is the product. Merge/split/sign/OCR are someone else’s toolbox.

MIT license.