[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"pack-detail-personal-knowledge-base-rag-zh":3,"seo:pack:personal-knowledge-base-rag:zh":102},{"code":4,"message":5,"data":6},200,"操作成功",{"pack":7},{"slug":8,"icon":9,"tone":10,"status":11,"status_label":12,"title":13,"description":14,"items":15,"install_cmd":101},"personal-knowledge-base-rag","🧠","#7C3AED","new","本周新建","个人知识库 — 让 AI 读你的笔记、日记、PDF","十件开源资产，把 AI 接到你自己的笔记、日记和 PDF 归档上 — 全程私有。笔记 app、MCP\u002Fagent 桥、本地索引器、纸质文档 OCR、文献管理。真正属于你的 second brain。",[16,28,38,48,55,65,72,79,87,94],{"id":17,"uuid":18,"slug":19,"title":20,"description":21,"author_name":22,"view_count":23,"vote_count":24,"lang_type":25,"type":26,"type_label":27},1112,"42404795-364b-11f1-9bc6-00163e2b0d79","logseq-privacy-first-knowledge-management-platform-42404795","Logseq — Privacy-First Knowledge Management Platform","Logseq is a privacy-first, open-source platform for knowledge management and collaboration. Outliner-based with bidirectional links, block references, queries, and graph visualization. Local-first with optional sync. The open-source alternative to Roam Research.","Script Depot",232,0,"en","script","Script",{"id":29,"uuid":30,"slug":31,"title":32,"description":33,"author_name":34,"view_count":35,"vote_count":24,"lang_type":25,"type":36,"type_label":37},3834,"f1cfca7b-2432-5ee2-a097-8dfe725140d9","obsidian-agent-client-bring-agents-into-obsidian","Obsidian Agent Client — Bring Agents into Obsidian","Obsidian Agent Client adds ACP-based agent integrations (Claude Code\u002FCodex\u002FGemini) to Obsidian; install via BRAT or Releases and use adapters (2,016★).","Agent Toolkit",86,"agent","Agent",{"id":39,"uuid":40,"slug":41,"title":42,"description":43,"author_name":44,"view_count":45,"vote_count":24,"lang_type":25,"type":46,"type_label":47},3242,"d245970b-97c3-4255-bc2a-ce4765ea3c75","obsidian-mcp-tools-notes-to-tools-for-claude","Obsidian MCP Tools — Notes-to-Tools for Claude","Obsidian MCP Tools adds a local MCP server via an Obsidian plugin, so Claude Desktop can query and automate your vault with a signed server component.","MCP Hub",49,"mcp","MCP",{"id":49,"uuid":50,"slug":51,"title":52,"description":53,"author_name":44,"view_count":54,"vote_count":24,"lang_type":25,"type":46,"type_label":47},3563,"ad293269-09bf-5104-bfb5-eb67a696c92f","joplin-mcp-mcp-server-for-joplin-notes","joplin-mcp — MCP Server for Joplin Notes","joplin-mcp is an MCP server for Joplin that lets agents read\u002Fsearch\u002Fedit notes; includes an installer for Claude Desktop and a Claude Code plugin.",36,{"id":56,"uuid":57,"slug":58,"title":59,"description":60,"author_name":61,"view_count":62,"vote_count":24,"lang_type":25,"type":63,"type_label":64},2820,"e8255b25-1bb1-47a8-bff9-ca5a445ce3f1","cherry-studio-knowledge-base-local-rag-with-50-formats","Cherry Studio Knowledge Base — Local RAG with 50+ Formats","Cherry Studio Knowledge Base ingests PDFs, Office docs, Markdown into a local vector index. Query offline, BYOK any LLM. Data stays on your machine.","Cherry Studio",118,"skill","Skill",{"id":66,"uuid":67,"slug":68,"title":69,"description":70,"author_name":22,"view_count":71,"vote_count":24,"lang_type":25,"type":26,"type_label":27},1616,"677aaba2-39c1-11f1-9bc6-00163e2b0d79","trilium-notes-self-hosted-hierarchical-knowledge-base-677aaba2","Trilium Notes — Self-Hosted Hierarchical Knowledge Base","Trilium Notes is a self-hosted, hierarchical note-taking application with rich editing, note cloning, relation maps, and scripting capabilities.",105,{"id":73,"uuid":74,"slug":75,"title":76,"description":77,"author_name":22,"view_count":78,"vote_count":24,"lang_type":25,"type":63,"type_label":64},2102,"44641b25-4080-11f1-9bc6-00163e2b0d79","blinko-self-hosted-personal-ai-note-taking-tool-44641b25","Blinko — Self-Hosted Personal AI Note-Taking Tool","An open-source, self-hosted personal note-taking tool with AI-powered tagging, search, and organization, built with Next.js and PostgreSQL.",129,{"id":80,"uuid":81,"slug":82,"title":83,"description":84,"author_name":85,"view_count":86,"vote_count":24,"lang_type":25,"type":63,"type_label":64},4612,"74dca4cd-5468-11f1-9bc6-00163e2b0d79","zotero-free-research-source-manager-citation-tool-74dca4cd","Zotero — Free Research Source Manager and Citation Tool","Zotero is a free, open-source reference management tool that helps you collect, organize, annotate, cite, and share research sources. Available on Windows, macOS, Linux, and iOS, it supports one-click saving from browsers and generates citations in thousands of styles.","AI Open Source",32,{"id":88,"uuid":89,"slug":90,"title":91,"description":92,"author_name":22,"view_count":93,"vote_count":24,"lang_type":25,"type":63,"type_label":64},926,"de0041a5-34b7-11f1-9bc6-00163e2b0d79","paperless-ngx-self-hosted-document-management-ocr-de0041a5","Paperless-ngx — Self-Hosted Document Management with OCR","Paperless-ngx is an open-source document management system that scans, OCRs, indexes, and archives all your physical and digital documents for full-text search.",245,{"id":95,"uuid":96,"slug":97,"title":98,"description":99,"author_name":22,"view_count":100,"vote_count":24,"lang_type":25,"type":63,"type_label":64},885,"d4bf6827-349d-11f1-9bc6-00163e2b0d79","memos-open-source-self-hosted-note-taking-tool-d4bf6827","Memos — Open Source Self-Hosted Note-Taking Tool","Memos is a lightweight, self-hosted note-taking app with Markdown support, tag-based organization, and a Twitter-like timeline interface for quick idea capture.",94,"tokrepo install pack\u002Fpersonal-knowledge-base-rag",{"pageType":103,"pageKey":8,"locale":104,"title":105,"metaDescription":106,"h1":107,"tldr":108,"bodyMarkdown":109,"faq":110,"schema":126,"internalLinks":133,"citations":146,"wordCount":159,"generatedAt":160},"pack","zh","个人知识库 — 10 个开源工具让 AI 读你的笔记、日记、PDF","Logseq \u002F Obsidian Agent Client \u002F Obsidian MCP Tools \u002F joplin-mcp \u002F Cherry Studio 知识库 \u002F Trilium \u002F Blinko \u002F Zotero \u002F Paperless-ngx \u002F Memos — 一套完整的个人 RAG 栈，把 AI 接到自己的笔记、日记和 PDF 归档上，不上传任何东西到云。","个人知识库 — 一套属于自己的私有 RAG 栈","十个开源工具，给「自己的人生」做 RAG：明文 Markdown 的笔记 app、把 AI 接进来的 MCP \u002F agent 桥、能吃 PDF 和纸质扫描件的本地索引器，再加一个文献管理。真正属于自己的 second brain。","## 这个 pack 包含什么\n\n这是当你想让 AI 读**你自己**的笔记 —— 日记、会议记录、写一半的随笔、读研时囤的 PDF —— 并回答相关问题，又不想把任何东西传到云端时，会搭的那套栈。\n\n这跟「找前沿模型聊天」是两件事。那边要的是推理能力；这边要的是**对你私有语料的召回**。你不需要 GPT-5 —— 你需要管道：一个把内容存成明文 Markdown 的笔记 app、一个能读你历史上所有格式的索引器、一个让 AI 查到笔记的 MCP 或 agent 桥、再加一个聊天前端。\n\n这个 pack **不等于** TokRepo 上已有的 `local-first-ai` pack —— 那一套搭的是完整私有 AI 工作流（Ollama、Open WebUI、ComfyUI、出图、转写）。这一套**只聚焦在个人 RAG 这一层**，把它做深：笔记存哪、怎么被索引、agent 怎么读到。\n\n## 推荐安装顺序\n\n1. **Logseq** — 大纲式、隐私优先的笔记 app。本地 Markdown 文件、可选加密同步、日记是一等公民。如果还没有信得过的笔记 app，从这里开始。不选 Notion 或 Bear 的原因：每条笔记都是磁盘上明文文件，下游任何工具都能读。\n2. **Obsidian Agent Client** — 已经是 Obsidian 用户的话，这个插件把 ACP 协议的 agent（Claude Code、Codex、Gemini）直接接进编辑器。你的 vault 就是 agent 的上下文。通过 BRAT 或 releases 安装。\n3. **Obsidian MCP Tools** — Obsidian 的另一条路：不是把 agent 拉进 Obsidian，而是把 Obsidian 暴露给 Claude Desktop —— 一个本地 MCP server。Claude 现在能查 vault、搜笔记、跑 dataview 查询。如果 Claude Desktop 是你日常主力，选这个。\n4. **joplin-mcp** — 给 Joplin 用户的同款方案。MCP server，让任何兼容 Claude Desktop 的客户端读写你的 Joplin 笔记。Joplin 可选 E2EE + 本地 MCP，笔记完全不出本机。\n5. **Cherry Studio 知识库** — 一站式选项。带本地 RAG 引擎的桌面 AI 客户端，能吃 50+ 格式（PDF、Markdown、DOCX、EPUB、网页、甚至 Notion 导出），本地 embedding，再喂给你连的任何 LLM。想用一个 app 而不是五个，从这里开始。\n6. **Trilium Notes** — 自托管的层级 KB，带脚本、属性搜索、真正的树形结构。Logseq 的平铺图谱模型走到头之后接它。如果你的脑子按大纲组织，又想每条笔记都能作为结构化数据被查询，是对的选择。\n7. **Blinko** — 自托管的个人 AI 笔记，RAG 已经内置。不用自己接索引器，开箱即用。更接近「私有 NotebookLM」而不是「Obsidian + 插件」。代价：可定制性弱，但开箱可用。\n8. **Zotero** — 文献管理器。是你的笔记和学术 \u002F PDF 世界之间的桥。自动从论文抓元数据、建可搜索的文献库、暴露本地 API，下游 RAG 工具（Cherry Studio、自写脚本）能索引它。\n9. **Paperless-ngx** — 带 OCR 的自托管文档管理。大多数知识库方案漏掉的一块：每张纸质账单、合同、收据扫描、OCR、打标、索引。然后 AI 才能回答「我 2024 年 3 月的电费是多少」 —— 你一个 PDF 都不用翻。\n10. **Memos** — 轻量自托管笔记。Twitter 式短捕捉，带标签可搜索。微博体日记层 —— 捕捉那些小到不该进 Logseq 但又重要到不能丢的想法。\n\n## 它们怎么协同\n\n```\n  ┌────────────────────────────────────────────────────────┐\n  │  你的私有语料（一切落盘，不上传）                      │\n  └────────────────────────────────────────────────────────┘\n     │           │              │              │\n  Logseq      Obsidian        Joplin         Trilium\n  Memos       （或一站式 Blinko）\n     │           │              │              │\n     │           ├─ Obsidian Agent Client（编辑器内 ACP）\n     │           ├─ Obsidian MCP Tools  ──► Claude Desktop\n     │                          │\n     │                          └─ joplin-mcp ──► Claude Desktop\n     │\n     └──► Cherry Studio 知识库  ──┐\n                                  │\n  Zotero（论文）  ─────────────► ├──► 本地 LLM \u002F Claude \u002F GPT\n  Paperless-ngx（扫描 PDF）────► │     （你自己选模型）\n                                  ▼\n                            跟知识库对话\n```\n\n模式：**笔记 app** 存原文，**桥层**（MCP server、agent 插件、内置 RAG 引擎）让它可查，**LLM 客户端**负责问。不用全装 —— 选跟你已有的笔记 app 对应的那一行，再按需加 Zotero 和 Paperless-ngx，如果你的知识不全在 Markdown 里。\n\n## 你会遇到的取舍\n\n- **MCP 桥 vs app 内 agent vs 一站式** —— 三种架构都成立。**MCP 桥**（Obsidian MCP Tools、joplin-mcp）保留你现有的笔记 app，让 Claude Desktop 伸手进来；适合有固定客户端的重度用户。**app 内 agent**（Obsidian Agent Client、Blinko）把 AI 塞进编辑器；适合想让答案紧贴原文的人。**一站式**（Cherry Studio、Blinko）把索引器 + 聊天 + 模型打包；适合不想维护三个工具的人。\n- **embedding 质量 vs 配置痛苦** —— 省事路径：用笔记 app 自带的索引器（Blinko、Cherry Studio）。高质量路径：通过 Ollama 跑 BGE-M3 或 nomic-embed-text，再把自己写的 RAG 管道指向 vault。多数人高估了召回质量对个人笔记的重要性 —— 语料体量小，你大概也记得答案在哪一带。先简单做。\n- **聊天层用前沿模型 vs 本地模型** —— RAG + Claude 4.5 \u002F GPT-5 综合质量最好。RAG + 本地 Llama 3.1 8B 日记 100% 私有，但综合能力明显弱。混合方案没问题：检索完全本地（笔记的 embedding 永远不出机器），只把 top-3 chunk 加你的问题发给前沿模型。\n- **PDF 是隐藏杀手** —— Markdown 笔记几秒就索引完。扫描 PDF 要 OCR（Paperless-ngx），学术 PDF 要懂版式的抽取（Zotero 管引用元数据；要做全文 RAG 可能还得在前面加 GROBID 或 unstructured.io）。任何非文本源都要单独规划一遍处理。\n\n## 常见踩坑\n\n- **一上来就全索引** —— 一万个文件的 vault 配上烂的 chunking 策略，给你的是一万段没用的 4 行片段。从**一个**子文件夹开始（比如今年的会议记录），测一下召回质量，再扩展。\n- **日记噪音淹没真正的知识** —— 如果每天又往同一个 vault 灌 2 KB 站会笔记，召回结果天天是昨天的待办。把语料分开：日常日记放 Logseq，长期笔记放 Obsidian，只索引长期 vault。\n- **MCP server 要权限太狠** —— MCP 给 AI 的能力很大。装之前先读每个 server 的权限清单。joplin-mcp 和 Obsidian MCP Tools 默认都是读 + 写；先用**只读**模式跑通流程再开写。\n- **烂扫描的 OCR 质量** —— Paperless-ngx 不错但不是魔法。手机斜着拍的 30 度收据回来是乱码。要真正可搜的东西用平板扫描仪，或 iOS Notes 的扫描功能（自动校正）。\n- **忘了备份 vault** —— 这整套的意义就是这是你的 second brain。加密 + 备份到自己存储（Syncthing、私有 rsync target、Joplin 自带服务器）。十年日记别全压在一台笔记本上。",[111,114,117,120,123],{"q":112,"a":113},"我已经在用 Obsidian — 这套里哪些是真需要的？","Obsidian 是你的根据地的话，pack 里有两条路，都收了。路径 A：装 Obsidian Agent Client，把 agent 接进编辑器 — 适合让 AI 的回答紧挨着你正在写的笔记。路径 B：装 Obsidian MCP Tools，让 Claude Desktop（或任何 MCP 客户端）从外面查你的 vault — 适合想要独立的聊天面板，vault 只是知识源。重度用户最后通常两个都跑。pack 里其他笔记 app 跳过 — 那些是给还没选定的人的。",{"q":115,"a":116},"聊天层用 Claude 或 GPT，这还算真隐私吗？","部分算。本地索引器（Cherry Studio 的 RAG 引擎、MCP server、笔记文件系统）让全部语料留在盘上。提问时只有 top-K 召回的 chunk 加你的 prompt 发给模型方。比起把整个 vault 上传到 ChatGPT，数据量降了几个数量级 — 但不是零。要全私有，把聊天层走 Ollama 的本地模型（参考 local-first-ai pack）。现实折中：用 Claude\u002FGPT 做综合，但永远别把真敏感内容（医疗、法律、隐私）放进被索引的文件夹。",{"q":118,"a":119},"这套跟已有的 local-first-ai pack 区别在哪？","local-first-ai 是完整个人 AI 套件：聊天（Open WebUI）、代码（Continue）、出图（ComfyUI）、转写（Faster Whisper）+ Khoj 和 Joplin 作为知识层。这个 pack 是「个人 RAG」这一层的窄切片，往深做：多种笔记 app 选项（Logseq \u002F Obsidian \u002F Trilium \u002F Blinko \u002F Memos）、让 Claude Desktop 读 vault 的 MCP 桥（Obsidian MCP Tools \u002F joplin-mcp）、纸质文档扫描（Paperless-ngx）、学术研究（Zotero）。没有 model runner —— 模型从 local-first-ai 借，或者让 Cherry Studio 接任何 API。",{"q":121,"a":122},"能用在 PDF 和纸质扫描件上吗，不只是 Markdown？","可以，所以 Paperless-ngx 和 Zotero 才进了 pack。Paperless-ngx 对扫描的收据、合同、账单、税务文档跑 OCR，再暴露可搜的索引。Zotero 处理学术 PDF，抽取元数据、存全文。Cherry Studio 知识库可以直接吃这两种格式。再冷门点的格式（EPUB \u002F DOCX \u002F 网页归档）Cherry Studio 开箱支持 50+ 种。模式：任何格式最终变成文本、文本变成 embedding、embedding 变成可搜。PDF 只是最慢的那一步。",{"q":124,"a":125},"不想装十个工具的话，最小可用配置是什么？","三个工具，按顺序：(1) 你真会每天用的笔记 app —— 从头开始选 Logseq，否则继续用 Obsidian 或 Joplin；(2) 跟它对应的桥层 —— Obsidian 选 Obsidian Agent Client 或 Obsidian MCP Tools、Joplin 选 joplin-mcp、想要一个 app 干所有事就直接 Cherry Studio；(3) 有一堆要可搜的纸质文档就加 Paperless-ngx。这是最小可跑的个人 RAG。研究者再加 Zotero、喜欢微博式捕捉再加 Memos、笔记 app 用到吃力再考虑换 Trilium 或 Blinko。",{"@context":127,"@type":128,"name":129,"description":130,"numberOfItems":131,"inLanguage":132},"https:\u002F\u002Fschema.org","ItemList","个人知识库 — RAG over Your Notes","十个开源工具搭一套私有 RAG 栈，对你自己的笔记、日记和 PDF 归档做检索 —— 笔记 app、MCP 桥、本地 RAG 引擎、纸质文档 OCR、文献管理。",10,"zh-CN",[134,138,142],{"url":135,"anchor":136,"reason":137},"\u002Fzh\u002Flocal-first-ai","本地优先 AI — 完整私有 AI 套件","互补 pack：这一套是知识层，那一套是聊天 \u002F 代码 \u002F 出图 \u002F 转写",{"url":139,"anchor":140,"reason":141},"\u002Fzh\u002Fai-memory","AI 记忆层与个人知识库","更宽的话题页，覆盖 Mem0 \u002F Zep 和端上记忆替代方案",{"url":143,"anchor":144,"reason":145},"\u002Fzh\u002Ffeatured","TokRepo 精选资产","这十个工具属于更大的精选目录",[147,151,155],{"claim":148,"source_name":149,"source_url":150},"Logseq 是隐私优先的开源知识管理平台，本地 Markdown 存储","Logseq 官网","https:\u002F\u002Flogseq.com\u002F",{"claim":152,"source_name":153,"source_url":154},"Model Context Protocol (MCP) 是开放标准，让桌面 AI 客户端连接到本地数据源","Anthropic MCP 公告","https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fmodel-context-protocol",{"claim":156,"source_name":157,"source_url":158},"Paperless-ngx 是带 OCR 的自托管文档管理系统","Paperless-ngx 项目文档","https:\u002F\u002Fdocs.paperless-ngx.com\u002F",880,"2026-05-22T00:00:00Z"]