AI 提炼原子笔记,过滤信息垃圾,把阅读转化为可检索的知识节点。
原子笔记(Atomic Note)是 Obsidian 核心理念之一——每条笔记只记录一个知识点,短小精悍、独立可读、可复用。
本插件帮你把长文、网页、选中文本,用 AI 一键提炼成规范的原子笔记,自动去重后存入你的知识库。
插件采用七阶段流水线处理,从原始输入到最终保存,每一步都有质量把关:
| 阶段 | 名称 | 说明 |
|---|---|---|
| Phase 1 | 读取内容 | 从文本、URL 或剪贴板获取原始内容 |
| Phase 2 | 质量门控 | 多维规则前置过滤低质/噪声内容(硬阻断 + 软警告 + 强制提炼) |
| Phase 3 | AI 提炼 | 调用 DeepSeek 将内容拆解为原子笔记 |
| Phase 4 | 同批去重 | BM25 + 中文分词 + 综合加权评分(余弦 0.5 + 关键词 0.3 + 标题 0.2) |
| Phase 4b | 知识库去重 | SimHash 预过滤 + BM25 余弦,与已有笔记高效比对 |
| Phase 5 | 内容核查 | 三层管线:原文溯源 → 语义比对 → 超源标记,核查事实声明和数据准确性 |
| Phase 6 | 笔记复查 | 洞见 + 知识直加(2-10),四级制(差/中/良/优),低于策略门槛自动过滤 |
最终输出经过质量筛选的原子笔记,可预览确认或自动保存至指定文件夹。
采用 BM25 + 中文分词 + 余弦相似度,配合 SimHash 预过滤和综合加权评分:
相比传统 TF-IDF:BM25 避免高频术语污染向量,jieba 风格分词提高跨表述匹配能力,SimHash 大幅降低计算量。
从每条笔记中提取事实声明(数字、百分比、日期、实体名称),通过三层管线逐条核查:
标记结果:已溯源 / 需对比 / 超源
AI 从两个维度对每条笔记打分(各 1-5 分):
总分 = 洞见 + 知识(2-10),分四级:差(2-3) 中(4-5) 良(6-7) 优(8-10)。低于策略门槛的笔记被自动过滤,不进入知识库。这是提炼后的最后一道质量防线。
Bamboo Darts: 从选中文本提炼原子笔记Bamboo Darts: 从 URL 提炼原子笔记Bamboo Darts: 从剪贴板提炼原子笔记Bamboo Darts: 打开面板 - 右侧栏Bamboo Darts: 打开面板 - 左侧栏Bamboo Darts: 打开面板 - 新标签页Bamboo Darts: 打开面板 - 分屏在编辑器中选中文本后右键,点击"提炼原子笔记"
点击左侧边栏的 ⚛️(atom)图标
在 Obsidian 设置 → Bamboo Darts 中配置:
| 配置项 | 说明 | 默认值 |
|---|---|---|
| API Key | 你的 DeepSeek API Key(必需)— AES-256 加密存储,换设备需重填 | — |
| API URL | DeepSeek API 地址 | https://api.deepseek.com/v1/chat/completions |
| 模型 | 使用的 DeepSeek 模型 | deepseek-v4-flash |
| 最大 Token 数 | AI 输出的最大 Token 数 | 6000 |
| 目标文件夹 | 原子笔记保存的文件夹 | 原子笔记 |
| 文件名模板 | 支持变量 {{title}}, {{date}}, {{time}}, {{timestamp}} |
{{title}} |
| 自动保存 | 开启后,提炼完成后仍展示结果弹窗,但默认全选所有笔记 | 关闭 |
| 去重目标文件夹 | 去重比对的专用文件夹,留空则复用"目标文件夹" | 留空 |
| 标签词汇表 | 偏好标签,逗号或换行分隔 | — |
| 标签模式 | 宽松:优先使用偏好标签,允许新增;严格:仅使用偏好标签 | 宽松 |
| 自动创建反向链接 | 从选中文本提炼时,在源文件插入笔记链接 | 关闭 |
| 启用内容核查 | 提炼后自动核查事实声明和数据准确性(Phase 5) | 开启 |
| 仅保存可溯源笔记 | 开启时自动取消存疑/无据笔记的复选(需先启用内容核查) | 关闭 |
| 启用笔记复查 | AI 二次评分,自动过滤低质量笔记(Phase 6) | 关闭 |
| 复查模型(可选) | 复查用模型,留空则复用提炼模型 | — |
| 复查 API URL(可选) | 复查用 API 地址,留空则复用提炼 API 地址 | — |
| 复查 API Key(可选) | 复查用 API Key,留空则复用提炼 API Key | — |
| 启用关联推荐 | 选中笔记后显示 Top10 相关笔记 | 开启 |
| 启用发现索引 | 缓存笔记特征加速发现 Tab,几千篇笔记也能秒级响应 | 开启 |
| 索引最大笔记数 | 发现索引最多缓存多少篇笔记的特征(0=不限制) | 500 |
| Jaccard 相似度门槛 | 候选笔记与当前笔记的关键词重叠度最低要求(0=不过滤) | 0 |
| MMR 相关度权重 | 推荐结果的相关度 vs 多样性平衡(0=纯多样性,1=纯相关度) | 0.7 |
| 推荐结果数量 | 发现 Tab 显示的推荐笔记数量 | 10 |
| 智能识别文章类型 | 自动判断内容特征,选择最合适的过滤策略 | 开启 |
| 过滤策略 | 手动指定过滤强度(技术文献 / 通用文章 / 观点评论) | — |
| 高级参数调整 | 手动调整各策略的去重阈值和质量门槛 | — |
| 启用深度提炼模式 | 对超长文章自动分段提炼,消耗更多 token | 关闭 |
| 输入截断长度 | 送入 AI 前截断原文的最大字符数 | 10000 |
| 面板位置 | 插件面板显示位置(右侧栏 / 左侧栏 / 新标签页 / 分屏) | 右侧栏 |
在 Obsidian 设置 → 社区插件中搜索 Bamboo Darts 安装。
miaoziguan/obsidian-Bamboo-Darts.obsidian/plugins/ 目录插件界面包括:命令面板(Command Palette)、提炼结果弹窗(Result Modal)、设置页面(Settings Tab)。
Q:是否需要付费 API?
A:需要 DeepSeek API Key,DeepSeek 有免费额度,具体请参考 DeepSeek 官网。
Q:支持离线使用吗?
A:不支持,本插件依赖 DeepSeek API 进行内容提炼。
Q:笔记保存到哪里?
A:默认保存到 原子笔记 文件夹,可在设置中自定义。
Q:API Key 安全吗?
A:API Key 使用 AES-256-GCM 加密后存储在本地,密钥由机器指纹派生(平台 + 主机名 + 用户名),同步到其他设备无法解密。
MIT
Atomic Notes are a core concept in Obsidian—each note captures exactly one knowledge point: concise, self-contained, and reusable.
This plugin helps you transform long articles, web pages, or selected text into well-structured atomic notes using AI, with automatic deduplication before saving to your vault.
The plugin uses a 7-stage pipeline, with quality checks at each step:
| Phase | Name | Description |
|---|---|---|
| Phase 1 | Read Content | Fetch raw content from text, URL, or clipboard |
| Phase 2 | Quality Gate | Multi-dimensional rules filter low-quality/noisy content (hard block + soft warning + forced extraction) |
| Phase 3 | AI Extraction | Call DeepSeek API to decompose content into atomic notes |
| Phase 4 | Batch Dedup | BM25 + Chinese word segmentation + weighted combined score (cosine 0.5 + keyword 0.3 + title 0.2) |
| Phase 4b | Vault Dedup | SimHash 64-bit fingerprint pre-filter + BM25 cosine against existing notes |
| Phase 5 | Content Verification | Three-layer pipeline: source tracing → semantic compare → out-of-scope marking; verify factual claims and numeric data |
| Phase 6 | Note Review | AI re-scores notes from two dimensions (insight + knowledge value) to filter low-value output |
Final output: quality-filtered atomic notes, ready for preview or auto-save.
Uses BM25 + Chinese word segmentation + cosine similarity, with SimHash pre-filtering and weighted combined scoring:
Compared to traditional TF-IDF: BM25 prevents high-frequency term pollution, jieba-style segmentation improves cross-expression matching, and SimHash dramatically reduces computation.
Extract fact claims containing numbers, percentages, dates, and entity names from each note, and verify through a three-layer pipeline:
Results: Traced / Compare / Out-of-scope
AI scores each note from two dimensions (1-5 points):
Total = insight + knowledge (2-10), graded: Poor(2-3) Fair(4-5) Good(6-7) Excellent(8-10). Notes below the strategy threshold are automatically filtered out. This is the final quality checkpoint.
Bamboo Darts: Extract atomic notes from selected textBamboo Darts: Extract atomic notes from URLBamboo Darts: Extract atomic notes from clipboardBamboo Darts: Open Panel - Right SidebarBamboo Darts: Open Panel - Left SidebarBamboo Darts: Open Panel - New TabBamboo Darts: Open Panel - SplitRight-click on selected text in the editor, then click "Extract atomic notes"
Click the ⚛️ (atom) icon in the left sidebar
Configure in Obsidian Settings → Bamboo Darts:
| Setting | Description | Default |
|---|---|---|
| API Key | Your DeepSeek API Key (required) — AES-256 encrypted, re-enter on new devices | — |
| API URL | DeepSeek API endpoint | https://api.deepseek.com/v1/chat/completions |
| Model | DeepSeek model to use | deepseek-v4-flash |
| Max Tokens | Maximum tokens for AI output | 6000 |
| Target Folder | Folder for saving atomic notes | 原子笔记 |
| File Name Template | Supports {{title}}, {{date}}, {{time}}, {{timestamp}} |
{{title}} |
| Auto Save | When enabled, shows result modal with all notes pre-selected for review | Off |
| Tag Vocabulary | Preferred tags, separated by commas or newlines | — |
| Tag Mode | Loose: prefer preferred tags but allow new ones; Strict: only use preferred tags | Loose |
| Auto Create Backlinks | Insert note links in source file when extracting from selected text | Off |
| Dedup Target Folder | Separate folder for dedup comparison; leave empty to reuse "Target Folder" | Empty |
| Enable Content Verification | Auto-verify factual claims and numeric data after extraction (Phase 5) | On |
| Verified Only | Auto-uncheck questionable/unsupported notes (requires Content Verification enabled) | Off |
| Enable Note Review | AI re-scores notes and filters low-quality ones (Phase 6) | Off |
| Review Model (Optional) | Model for review, leave empty to reuse extraction model | — |
| Review API URL (Optional) | API endpoint for review, leave empty to reuse extraction API URL | — |
| Review API Key (Optional) | API Key for review, leave empty to reuse extraction API Key | — |
| Enable Related Recommendation | Show Top10 related notes when selecting a note | On |
| Enable Discovery Index | Cache note features to speed up discovery tab, works well with thousands of notes | On |
| Max Notes in Index | Maximum number of notes to cache in discovery index (0 = no limit) | 500 |
| Jaccard Similarity Threshold | Minimum keyword overlap required between candidate and current note (0 = no filter) | 0 |
| MMR Relevance Weight | Balance between relevance and diversity for recommendations (0 = pure diversity, 1 = pure relevance) | 0.7 |
| Recommendation Count | Number of recommended notes to show in discovery tab | 10 |
| Auto-classify Content Type | Automatically detect content type and select the best filter strategy | On |
| Filter Strategy | Manually specify filter intensity (technical / general / opinion) | — |
| Advanced Parameters | Manually adjust dedup thresholds and quality thresholds for each strategy | — |
| Enable Deep Extraction | Auto-chunk very long articles for extraction (uses more tokens) | Off |
| Input Truncation Length | Maximum characters of source text sent to AI | 10000 |
| Panel Position | Where the plugin panel appears in the Obsidian UI | Right sidebar |
Search for Bamboo Darts in Obsidian Settings → Community Plugins.
miaoziguan/obsidian-Bamboo-Darts.obsidian/plugins/ in your vaultQ: Is a paid API required?
A: A DeepSeek API Key is required. DeepSeek offers free credits—see the DeepSeek website for details.
Q: Does it work offline?
A: No, this plugin relies on the DeepSeek API for content extraction.
Q: Where are notes saved?
A: Notes are saved to the 原子笔记 folder by default; you can customize this in settings.
Q: Is my API Key secure?
A: API Keys are encrypted with AES-256-GCM before storage, with a key derived from your machine fingerprint (platform + hostname + username). Sync'd keys cannot be decrypted on other devices.
See CHANGELOG or the Releases page.
MIT