📖 DeepRead — Make AI reading traceable
Website · Real outputs · English · 中文
Turn articles, books, PDFs, and document sets into claims you can trace back to evidence and source locations.
DeepRead is available in two compatible forms:
- Portable Agent Skill for Codex, Claude Code, and other Agent Skills-compatible tools. Zero runtime dependencies; the agent follows the evidence-first reading workflow with its own file and web tools.
- Host plugin package for DeepSeek Harness Web/headless and dsh-TUI, with a
deepreadtool, PDF extraction, optional persistence/jobs/Web route, batch comparison, cost preview, and HTML/XMind-compatible export. Its browser client is an optional Web-only entry.
Why it is different
| A typical summary | DeepRead |
|---|---|
| Compresses the topic | Extracts complete claims and the reasoning behind them |
| Blends source facts with model inference | Labels author intent, source facts, reasoned inference, and unverified content |
| Makes conclusions hard to check | Pairs important claims with evidence and page/paragraph locations |
| Stops at an answer | Adds knowledge maps, conflicts, limitations, and active-recall questions |
If the source does not support a claim, DeepRead says “source does not provide evidence” instead of filling the gap.
Quick start
Portable Agent Skill
npx skills@latest add xiehuan123/dsh-deepread
Then ask your agent:
Deep-read docs/architecture.pdf in knowledge-map mode.
For every important claim, show the supporting evidence and source location.
Full DeepSeek Harness plugin
dsh plugin --profile web add dsh-deepread
If pnpm reports ERR_PNPM_ADDING_TO_ROOT, retry with the profile workspace made explicit:
dsh plugin --profile web add -w dsh-deepread
Restart dsh web, then use the 📖 reading panel or call the deepread tool in chat.
See real outputs
These are complete reports generated from public articles, not hand-written mockups:
| Report | What DeepRead made visible |
|---|---|
deep · Claude Code token optimization |
Reconstructed the engineering chain from visibility to input, output, and retrieval-path compression; separated recommendations from project-authored benchmarks. |
map · Marketing-claim fact check |
Found that the article's “90%”, “¥1.28M salary”, and “¥2,000/day” claims had no source, sample, or baseline; marked each one unverified. |
deep · vivo Tauri architecture |
Connected architecture choices to reported size/performance evidence while preserving the article's untested assumptions and deployment limits. |
Browse all reproducible examples →
Features
| Capability | Details |
|---|---|
| 🎛️ Five modes | quick key takeaways · deep in-depth reading · map knowledge map · feynman Feynman technique (11-step loop + spaced repetition) · book whole-book reading (see the comparison below) |
| 🗺️ Knowledge-map mode | Core question / core conclusion / ten content categories (conclusion, sub-claim, mechanism, fact, data, case, hidden premise, objection, limitation, actionable advice) / every claim paired with evidence (unverifiable claims marked "no evidence provided in the original text") / key data table (value & unit, time range, sample, baseline, source, location) / eight relation labels (supports, refutes, causes, explains, depends on, exemplifies, contrasts, limits) / four confidence levels (author intent, original facts & data, reasonable inference, unverifiable) / Mermaid mindmap / XMind outline / 5 active-recall questions |
| 📥 Three inputs | WeChat article URLs (mp.weixin.qq.com stable links) · files (.txt/.md/.html/.pdf, PDF via a built-in pure-JS extractor with Chinese ToUnicode mapping, page markers, and object-stream/xref-stream support) · pasted text |
| 📤 Optional export | Displayed in-session by default; export accepts md / mm (FreeMind, importable by XMind) / html (editor-style web report with light/dark theme) / all, written to deepread-output/ in the workspace |
| 🎨 Browser UI | deepread tool result card (four-color confidence legend, collapsible sections) + a 📖 shortcut button next to the input area that opens a card-style reading panel (link/path/text + mode/export selection + reading focus + one-click start) |
| 🔀 Batch compare | Pass 2-10 documents via batch (url/path/text each) to get per-document summaries plus a cross-document report: comparison matrix, conflicts, complementarity, and synthesis |
| 📍 Citations | Reports carry page/paragraph provenance: arguments, quotes, and a dedicated citation table locate claims back to 【第N页】 markers in the source |
| 🧮 Cost preview | estimate: true previews token spend, model-call count, and expected time per mode without calling the model (CJK≈0.6 tok/char heuristic; rate/latency defaults are picked per model family and can be overridden explicitly) |
| 📚 Recently read | The Web panel keeps a local history of recent reads with one-click re-read (localStorage, no server round-trip) |
| ⏳ Progress transparency | Long reads / big PDFs / batches become official background jobs: the label states segment count and budget; the progress stream pushes 「精读第 3/20 段…」 line by line; job_output polls progress and the final report, job_kill cancels |
| 🔍 Parse progress | Full PDF extraction moves inside the background job and streams per page — 「解析 PDF 中… 42%（10/24 页）」 — after a fast sampling preflight decides length (no more silent wait before the background job appears); batches stream per document — 「解析第 2/5 篇… / 精读第 2/5 篇… / 完成第 2/5 篇」 plus 「跨篇对比汇总中…」 |
| 🧮 Panel budget | The Web panel shows per-mode token + time hints above the mode chips (e.g. 深度精读 (≈38k token · ≈8分钟)), instantly for pasted text; calibrated by real model speed; links/file paths are fetched and estimated by the Host through a same-origin API (POST /api/deepread/budget) and the panel's 🔍 budget-preflight button shows a one-line result (≈N chars · ≈X token · ≈Y min) right inside the panel — no chat round-trip, no table |
| ⚡ Fast preflight | estimate mode samples the first 2 PDF pages and extrapolates by page count, so big PDF budgets come back in milliseconds |
| 🎯 Self-calibration | Real token/s measured from every model call feeds a rolling average persisted in storage — estimates converge to your actual provider speed; cold-start defaults are per model family (DeepSeek/Kimi/Qwen ≈100-110 tok/s, Claude ≈70, GPT ≈90) |
Five modes compared
| Mode | Best for | Key output | Cost |
|---|---|---|---|
quick |
"What is this article about?" at a glance | One-line summary, core claim, argument structure, quotes, key concepts, critical questions | Single call, fastest |
deep (default) |
Reading one article carefully | Overview, core claim, argument structure (claim + evidence + verbatim quotes), argument flow, section highlights, quotes, key concepts, critical thinking | Long articles are auto-split, section-by-section + summary |
map |
Research, fact-checking before citing | Core question & conclusion, ten content categories, claim-evidence pairing, key data table (five elements), eight relations, four confidence levels, Mermaid mindmap, XMind outline, active-recall questions | Structured pipeline, multiple calls |
feynman |
Truly learning it and teaching it to others | 11-step loop: TOC → questions → per-chapter → claims/data/evidence → chapter mindmap → explain with the book closed → self-check gaps → correct against the source → merged mindmap → explain again → spaced review on days 1/3/7/14/30 | Longest output, most calls |
book |
Whole books / very long texts | Table of contents, chapter flow, a full-book summary assembled from per-part deep reads | Processed part by part |
One-line picker: in a hurry, quick; read one article thoroughly, deep; cite and fact-check, map; learn and remember, feynman; a whole book, book.
Installation
DeepRead 1.0.1 requires Node.js 22.19 or 24 and higher (^22.19 || >=24). The same npm package exposes the TypeScript Host entry at lib/types/index.js, the dsh-TUI Community Consensus v0.15 manifest at dsh-plugin.json, and an optional DeepSeek Harness Web client at lib/client.js.
Host compatibility
| Host | Node deepread tool | Web client