deep-read-summarize — turn papers, videos, books & web pages into structured Obsidian deep-read notes

[PensiveFei]

I built deep-read-summarize, a DSH plugin that does deep reading, not summarization. Give it one prompt — an arXiv paper, a Bilibili/YouTube video, a book (PDF/EPUB/MOBI), or a web page — and it produces a multi-thousand-word structured note: YAML frontmatter, one-line takeaway, section-by-section analysis, action items, and further reading. A short digest shows up right in the chat while the full note lands in your vault.

One prompt → a structured, metadata-rich note (here: the Chain-of-Thought paper, NeurIPS 2022).

What it does

• 📥 Multi-type input: papers (arXiv/PDF/HTML), videos (subtitle-first, auto-transcribe when no subtitles exist), books, web pages
• 🗂️ Obsidian-ready output: YAML frontmatter (title / aliases / author / year / source / tags / status), Dataview-friendly
• 🧩 MapReduce pipeline: fetch → chunk → N subagents read in parallel → merge + quality check
• ✅ Subagent output constrained by JSON Schema with auto-retry; key claims carry page/section/paragraph citations to reduce hallucination
• 🛟 Fails loud on config errors; degrades gracefully on content errors (a failed chunk is skipped and marked as a gap, never silently dropped)
• 🌐 Note language follows your preference (examples below are Chinese; English works too — the CoT note keeps bilingual aliases)

Install (v0.3.6, zero third-party dependencies, MIT)

npm install deep-read-summarize
# or: pnpm add ./deep-read-summarize-0.3.6.tgz in your dsh profile,
# add "deep-read-summarize" to dsh.profile.bundles, restart dsh web

Numbers

• 630+ downloads/week on npm
• 25 offline fixture tests: npm test → TOTAL: 25 passed, 0 failed
• Notes typically run 3,000–6,000+ words depending on source length

Same pipeline on a Bilibili video (8 min 38 s, no subtitles → auto-transcribed).

Actionable checklists, not just a summary.

The in-chat digest. This run also shows the fallback: fetch + transcription exceeded the 10-minute per-call wall clock, so the agent read the transcript already on disk and finished the note itself — and said so.

Known limits: very long videos can exceed the per-call wall clock (fallback shown above); DSH is developer preview, so workflow-tool semantics may change (see README compatibility notes).

Disclaimer: unofficial third-party plugin — not affiliated with or endorsed by DeepSeek.

Feedback welcome → repo · npm

---

中文速览：deep-read-summarize 是一个 DSH 精读插件——给它一个 prompt（论文 / 视频 / 书籍 / 网页），它把内容分块、交给多个子代理并行精读，合并成一篇带 YAML frontmatter 的结构化 Obsidian 笔记（一句话概括、逐节拆解、行动要点、延伸阅读），对话里同时给一份速览。视频无字幕自动转写；关键结论标注出处；内容解析失败会标记缺口而不是静默丢弃。npm install deep-read-summarize 即可安装（25 项离线 fixture 测试全绿，npm 周下载 630+）。欢迎试用并在 issue 区反馈。