Skip to content
BranchesTags
 
 

Repository files navigation

Paper Daily

Obsidian plugin — Daily arXiv + HuggingFace paper digest with AI summarization, interest keyword ranking, deep read, and backfill.

中文说明 | English

中文说明

简介

Paper Daily 是一个 Obsidian 插件，每天自动从 arXiv 和 HuggingFace Daily Papers 拉取你关注领域的最新论文，通过 LLM 分批打分（10 篇/批）后重排，生成结构化每日摘要，并支持对排名最高的论文进行 HTML 全文精读并写入独立笔记。支持任意日期范围回补，所有操作均有悬浮进度组件实时反馈。

适合人群：AI/ML 研究者、工程师，希望在 Obsidian 笔记库中持续追踪 arXiv 和 HuggingFace 最新进展。

核心功能

功能 说明
每日拉取 arXiv 按分类 + 关键词检索，过滤过去 N 小时内的新论文，自动去重
HuggingFace 源 抓取 HF 每日精选，HF 点赞数作为排名首要信号
LLM 批量打分 每 10 篇调用一次 LLM 打分，结果用于重排序
兴趣关键词 配置带权重的关注词，命中论文排名靠前，摘要中高亮显示
今日兴趣热度 每日摘要顶部以表格形式展示关键词热度排名
AI 摘要 LLM 生成结构化要点，标注方向、关键词命中
精选论文 对完成深度精读的论文生成独立卡片，含 wikilink
Deep Read（深度精读） 抓取排名最高 N 篇论文的 arxiv.org/html 全文，逐篇 LLM 分析，写入独立笔记
回补（Backfill） 日历选择器指定日期范围，补充生成历史每日摘要；与日常运行并行，各有独立悬浮组件
悬浮进度组件 实时显示 token 消耗（紧凑格式：12.3k），含停止按钮
容灾保障 网络或 LLM 报错时仍然落盘文件（含错误说明），产物不断档
日志轮转 runs.log 超过 10MB 自动轮转

输出文件结构

PaperDaily/
  inbox/
    2026-02-28.md           每日摘要 Markdown
  papers/
    2026-02-28.json         原始论文数据快照
  deep-read/
    2026-02-28/
      paper-title-deep-read-model.md   深度精读笔记（每篇独立文件）
  cache/
    state.json              运行状态
    seen_ids.json           去重记录
    runs.log                运行日志（10MB 上限自动轮转）

安装

方法一：直接复制文件（推荐）

  1. 前往本仓库 Releases 页面，下载最新版本的以下文件：

    • main.js
    • manifest.json
    • styles.css（如有）

  2. 在你的 Obsidian Vault 中创建插件目录（如不存在）：

    <你的 Vault>/.obsidian/plugins/paper-daily/

  3. 将上述文件复制到该目录：

    .obsidian/plugins/paper-daily/
├── main.js
├── manifest.json
└── styles.css

  4. 打开 Obsidian → 设置 → 第三方插件，关闭「安全模式」，然后启用 Paper Daily。

macOS 路径示例：~/Documents/MyVault/.obsidian/plugins/paper-daily/ Windows 路径示例：C:\Users\你的用户名\Documents\MyVault\.obsidian\plugins\paper-daily\

方法二：开发者本地构建

git clone https://github.com/your-username/paper-daily.git
cd paper-daily
npm install
npm run build

将 main.js 和 manifest.json 复制到 Vault 插件目录，或使用符号链接：

ln -s $(pwd) ~/path/to/your/vault/.obsidian/plugins/paper-daily

快速开始

  1. 安装并启用插件后，打开 设置 → Paper Daily
  2. 填入你的 LLM API Key，选择服务商（DeepSeek / OpenAI / Claude 等）
  3. 确认 arXiv 分类（默认 cs.AI, cs.LG, cs.CL）
  4. 按 Ctrl+P 打开命令面板，执行 Paper Daily: Run daily fetch & summarize now
  5. 生成的摘要位于 PaperDaily/inbox/YYYY-MM-DD.md

配置说明

打开 Obsidian 设置 → Paper Daily：

arXiv 拉取

配置项 默认值 说明
Categories cs.AI,cs.LG,cs.CL arXiv 分类，逗号分隔
Keywords 空 查询关键词，与分类 AND 组合；留空则只按分类查询
Interest Keywords 空 关注词（格式 keyword:weight），用于排序和高亮
Max Results Per Day 20 每日摘要最多展示的论文数
Time Window 72h 拉取过去 N 小时内的论文
Sort By submittedDate 按提交日期或最后更新日期排序

HuggingFace 源

配置项 默认值 说明
Enable HF Source 开启 抓取 HF 每日精选，点赞数合并到排名中
Lookback Days 3 今日无数据时（如周末）向前查找的天数
Dedup HF Papers 关闭 跳过已在历史摘要中出现过的 HF 精选

LLM 配置

支持的服务商（设置页一键切换）：

服务商 类型 说明
DeepSeek OpenAI Compatible 推荐，性价比高
OpenAI OpenAI Compatible GPT-4o / GPT-4o-mini
Claude Anthropic SDK claude-3-5-sonnet / claude-opus-4
Qwen / 通义 OpenAI Compatible 阿里云 DashScope
GLM / 智谱 OpenAI Compatible —
Moonshot / Kimi OpenAI Compatible —
MiniMax OpenAI Compatible —
Custom OpenAI Compatible 任意 OpenAI 格式接口

Deep Read（深度精读）

配置项 默认值 说明
Enable Deep Read 关闭 开启后精读功能激活
Papers to Fetch (topN) 10 每日抓取全文的篇数，范围 1–999
File Name Template 可配置 精读笔记文件名模板，支持变量占位符

精读笔记写入路径：PaperDaily/deep-read/YYYY-MM-DD/[paper-name]-deep-read-[model].md

命令

在命令面板（Ctrl+P）搜索 Paper Daily：

命令 说明
Run daily fetch & summarize now 立即拉取今日论文并生成摘要
Backfill daily summaries for date range 打开日历选择器，回补指定日期范围的每日摘要
Rebuild index from local cache 从本地缓存重建去重索引
Open settings 打开插件设置页

每日摘要结构

---
type: paper-daily
date: 2026-02-28
sources: [arxiv, huggingface]
categories: [cs.AI, cs.LG, cs.CL]
---

# Paper Daily — 2026-02-28

## 今日兴趣领域热度
| 关键词 | 命中数 | 总权重 |
|--------|--------|--------|
| kv cache | 4 | 8.0 |
| agent | 3 | 6.0 |
...

## 今日要点（AI 总结）
- ...

## 精选论文 / Featured Papers
[[paper-title-deep-read-model]]
...

## All Papers
| 标题 | 评分 | 方向 | 命中 | arXiv | HF | Deep Read |
|------|------|------|------|-------|-----|-----------|
...

调度时间

任务 默认时间 说明
每日拉取 08:30 插件内置调度，无需系统 cron

路线图

  • 周报 / 月报自动生成
  • BibTeX 导出
  • RSS 数据源实现
  • 自定义 API 数据源实现
  • RAG 检索历史论文
  • 侧边栏 UI 面板

技术栈

TypeScript · Obsidian API · esbuild · arXiv Atom API · HuggingFace Papers · @anthropic-ai/sdk

English

Overview

Paper Daily is an Obsidian plugin that automatically fetches the latest papers from arXiv and HuggingFace Daily Papers every day. Papers are LLM-batch-scored (10 per batch) and re-ranked by score, then presented as a structured daily digest. Top-ranked papers can be sent for full-text deep read: the plugin fetches the full HTML from arxiv.org/html, runs a per-paper LLM analysis, and writes each result as a standalone note. Date-range backfill is supported, with parallel floating progress widgets for concurrent runs.

Ideal for: AI/ML researchers and engineers who want a persistent, searchable paper feed inside their Obsidian vault.

Features

Feature Description
Daily arXiv fetch Search by category + keywords, filter to past N hours, deduplicate
HuggingFace source Fetch HF daily featured papers; upvotes are the primary ranking signal
LLM batch scoring Score papers in batches of 10 per LLM call, then re-rank by score
Interest keywords Weighted personal watchlist — hits boost ranking and are highlighted in the digest
Keyword hotness table Top of each digest shows a keyword heat table ranked by hit count and weight
AI digest LLM-generated structured summary with direction tags and keyword highlights
Featured Papers section Deep-read papers appear as wikilinked cards in the digest
Deep Read Fetch top-N papers' full HTML from arxiv.org/html, run per-paper LLM analysis, write standalone notes; topN default 10, range 1–999
Backfill Calendar date-picker modal for selecting a past date range; runs in parallel with the daily pipeline via separate stacked floating widgets
Floating progress widget Real-time token display (compact format: 12.3k) with a stop button
Fault-tolerant writes Network or LLM failures still produce output files with an error note
Log rotation runs.log is capped at 10 MB and rotated automatically

Installation

Option 1: Copy Files (Recommended)

  1. Go to the Releases page and download the latest:

    • main.js
    • manifest.json
    • styles.css (if present)

  2. Create the plugin folder in your vault (if it doesn't exist):

    <YourVault>/.obsidian/plugins/paper-daily/

  3. Copy the files into that folder:

    .obsidian/plugins/paper-daily/
├── main.js
├── manifest.json
└── styles.css

  4. In Obsidian: Settings → Community Plugins → disable Safe Mode → enable Paper Daily.

macOS: ~/Documents/MyVault/.obsidian/plugins/paper-daily/ Windows: C:\Users\YourName\Documents\MyVault\.obsidian\plugins\paper-daily\

Option 2: Build from Source

git clone https://github.com/your-username/paper-daily.git
cd paper-daily
npm install
npm run build

Copy main.js + manifest.json to the vault plugin folder, or symlink the repo:

ln -s $(pwd) ~/path/to/your/vault/.obsidian/plugins/paper-daily

Quick Start

  1. Install and enable the plugin, then open Settings → Paper Daily
  2. Enter your LLM API Key and select a provider (DeepSeek / OpenAI / Claude, etc.)
  3. Confirm arXiv Categories (default: cs.AI, cs.LG, cs.CL)
  4. Open the command palette (Ctrl+P) and run Paper Daily: Run daily fetch & summarize now
  5. Find your digest at PaperDaily/inbox/YYYY-MM-DD.md

Configuration

arXiv Fetch

Setting Default Description
Categories cs.AI,cs.LG,cs.CL Comma-separated arXiv categories
Keywords empty Query keywords, ANDed with categories; leave empty for category-only search
Interest Keywords empty Personal watchlist in keyword:weight format, used for ranking and highlight
Max Results Per Day 20 Maximum papers shown in the daily digest after ranking
Time Window 72h Fetch papers from the past N hours
Sort By submittedDate Sort by submission date or last updated date

HuggingFace Source

Setting Default Description
Enable HF Source on Fetch HF daily featured papers; upvotes merged into ranking
Lookback Days 3 Days to look back when today has no data (e.g. weekends)
Dedup HF Papers off Skip HF papers already seen in past digests

LLM Provider

One-click presets in settings:

Provider Type Notes
DeepSeek OpenAI Compatible Cost-effective, recommended
OpenAI OpenAI Compatible GPT-4o / GPT-4o-mini
Claude Anthropic SDK claude-3-5-sonnet / claude-opus-4
Qwen OpenAI Compatible Alibaba DashScope
GLM / Zhipu OpenAI Compatible —
Moonshot / Kimi OpenAI Compatible —
MiniMax OpenAI Compatible —
Custom OpenAI Compatible Any OpenAI-format endpoint

Deep Read

Setting Default Description
Enable Deep Read off Activate full-text deep read for top papers
Papers to Fetch (topN) 10 Number of papers to deep-read per day; range 1–999
File Name Template configurable Template for note filenames with variable placeholders

Deep-read notes are written to: PaperDaily/deep-read/YYYY-MM-DD/[paper-name]-deep-read-[model].md

Commands

Open the command palette (Ctrl+P) and search Paper Daily:

Command Description
Run daily fetch & summarize now Immediately fetch today's papers and generate a digest
Backfill daily summaries for date range Open the calendar date-picker modal and backfill past digests
Rebuild index from local cache Reload the dedup index from disk
Open settings Open the plugin settings tab

Daily Digest Structure

Each daily digest is written to PaperDaily/inbox/YYYY-MM-DD.md with the following sections, in order:

  1. Frontmatter (type, date, sources, categories, ...)
  2. # Paper Daily — YYYY-MM-DD
  3. ## 今日兴趣领域热度 — keyword hotness table (hit count + weighted score)
  4. ## 今日要点（AI 总结） — LLM-generated key takeaways
  5. ## 精选论文 / Featured Papers — wikilinks to deep-read notes
  6. ## All Papers — full ranked table with arXiv link, HF link, Deep Read link, score, one-line summary, and interest hits

Vault Output Layout

PaperDaily/
  inbox/
    2026-02-28.md             daily digest
  papers/
    2026-02-28.json           raw paper snapshot
  deep-read/
    2026-02-28/
      paper-title-deep-read-model.md   per-paper deep-read note
  cache/
    state.json                run state
    seen_ids.json             dedup store
    runs.log                  run log (10 MB cap, auto-rotated)

Scheduling

Task Default Time Notes
Daily fetch 08:30 Built-in scheduler — no system cron required

The scheduler uses a 60-second tick with last-run timestamp checks to avoid double-runs.

Roadmap

  • Weekly / monthly report generation
  • BibTeX export
  • RSS source implementation
  • Custom API source implementation
  • RAG over historical papers
  • Sidebar UI panel

Stack

TypeScript · Obsidian API · esbuild · arXiv Atom API · HuggingFace Papers · @anthropic-ai/sdk

License

MIT

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages