English ｜ 中文说明

An Obsidian plugin: chat with multiple LLM providers, search your notes with a local knowledge base (RAG), run agents that can read and edit notes, keep long-term memory, and roleplay with character cards.

Since v2.0.0 this plugin is free and open source (MIT). The paid activation system has been removed from the code and all old license keys are void. Users who paid before can open an issue to request a refund.

Author 3zh ｜ License MIT ｜ Repo https://github.com/zzzzzllllllaaaa/ai-chat-assistant ｜ Changelog

Install

From the community directory (recommended)

1. Obsidian → Settings → Community plugins → Browse
2. Search for AI Chat Assistant → Install → Enable

Manually, from a GitHub release

1. Download main.js, manifest.json and styles.css from the latest release
2. Copy them into <your vault>/.obsidian/plugins/ai-chat-assistant/
3. Restart Obsidian → Settings → Community plugins → enable AI Chat Assistant

Requires Obsidian 1.7.2+. Runs on desktop and mobile.

Getting started

1. Settings → AI Chat Assistant → Models → add a connection (name / base URL / API key) → test it → fetch the model list.
2. Bind models per purpose — chat, agent and embedding are configured separately (the knowledge base needs a working embedding model).
3. For retrieval, run Update knowledge base index (incremental) from the command palette.

With Ollama you can point the base URL at your local server and leave the API key empty to run fully offline.

Features

• Multi-provider chat — OpenAI, Claude, Gemini, DashScope (Qwen), DeepSeek or any OpenAI-compatible endpoint; local models through Ollama; streaming output; per-purpose model binding (chat / agent / embedding).
• Knowledge base (RAG) — hybrid search (vector similarity + keywords + recency) fused with RRF, query rewriting and reranking, graph-aware retrieval that follows your note links, vectors stored in IndexedDB with a Web Worker so the UI stays responsive.
• Agents — tool calling (read and write notes, create folders, full-text search, web search…), multi-step task execution with progress summaries, and permission prompts before a tool changes anything.
• Memory and roleplay — long-term memory, character cards (SillyTavern-compatible), multi-character group chat, writing styles.
• MCP tools — connect standard MCP servers over HTTP / SSE.
• Inline AI — rewrite or continue a selection from the editor.

Privacy and data

• No telemetry — the plugin contains no analytics or data collection code.
• Network use is limited to the endpoints you configure: LLM providers, MCP servers, web search, and the optional remote skill/service APIs.
• Chat content and notes are only sent to the model provider you configure. With a local model and the local vector store, the plugin can run fully offline.

Disclosure, so that nothing surprises you:

• Files outside the vault — the agent's note tools use the Obsidian vault API, but some features (skill installers, importing/exporting packages, the built-in HTTP API server) use Node's fs and can read or write paths outside your vault. Nothing is touched unless you trigger those features.
• Shell execution — certain skill features can run the scripts shipped inside a skill you installed, through child_process. Tool calls are gated behind permission prompts, but installing an untrusted skill is still equivalent to running its code.
• Clipboard — used for the copy/paste actions in chat.
• Dynamic code in skills — installed skills may be executed as scripts. Only install skills you trust.

Known limitations

• DeepSeek prompt caching is not properly optimised — the prompt prefix is not stable, so the cache benefit is much lower than expected.
• Skills and MCP are early work — MCP supports HTTP/SSE only (no stdio) and has not been exercised against many real-world servers.
• Large files, real debt — main.ts and src/ui/views/view.ts are several thousand lines each; the split is unfinished.
• No automated tests — changes are verified by hand.
• The built-in HTTP API server has no authentication (off by default, meant for localhost only — do not expose it to the internet).
• Mobile is supported but slow — indexing and retrieval on a large vault are not great.

License

MIT. Source code is public; releases carry the built artifacts.

Bug reports and ideas: GitHub Issues. This is a personal project and I make no promise about a maintenance schedule.

中文说明

Obsidian 插件：多模型对话 + 本地知识库检索（RAG）+ 能读写你笔记的智能体 + 长期记忆与角色扮演。

从 v2.0.0 起完全免费、源码公开（MIT），代码里已没有激活环节，所有旧激活码作废。曾付费的用户可以开 issue 联系退款。

作者 3zh ｜ 协议 MIT ｜ 仓库 https://github.com/zzzzzllllllaaaa/ai-chat-assistant ｜ 更新日志

安装

从社区目录装（推荐）

1. Obsidian → 设置 → 第三方插件 → 浏览
2. 搜索 AI Chat Assistant → 安装 → 启用

手动装（从 GitHub Release）

1. 到 最新 Release 下载 main.js、manifest.json、styles.css
2. 放进 <你的库>/.obsidian/plugins/ai-chat-assistant/
3. 重启 Obsidian → 设置 → 第三方插件 → 启用 AI Chat Assistant

需要 Obsidian 1.7.2+，桌面端与手机端都能跑。

快速开始

1. 设置 → AI Chat Assistant → 模型 → 添加连接（名称 / Base URL / API Key）→ 测试连接 → 拉取模型。
2. 按用途绑定模型 —— 聊天、智能体、向量是分开配的（用知识库必须配一个可用的向量模型）。
3. 要用检索：命令面板执行 更新知识库索引（增量）。

用 Ollama 就把 Base URL 填本地地址、Key 留空，可以完全离线运行。

功能

• 多模型对话 —— OpenAI、Claude、Gemini、通义千问（DashScope）、DeepSeek 或任何 OpenAI 兼容端点；本地模型走 Ollama；流式输出；按用途绑定不同模型（聊天 / 智能体 / 向量）。
• 知识库（RAG） —— 混合检索（向量相似度 + 关键词 + 时间衰减）用 RRF 融合，查询改写与结果重排，图谱感知（顺着笔记双链取上下文）；向量存 IndexedDB，检索跑在 Web Worker 里，不阻塞界面。
• 智能体 —— 工具调用（读写笔记、建文件夹、全文搜索、网络搜索等），多步任务执行带阶段小结，动手改东西前会弹权限确认。
• 记忆与角色扮演 —— 长期记忆、角色卡（兼容 SillyTavern）、多角色群聊、写作风格。
• MCP 工具 —— 接标准 MCP Server（HTTP / SSE）。
• 内联 AI —— 在编辑器里改写或续写选中内容。

隐私与数据

• 无遥测：插件不含埋点或数据收集代码。
• 网络只连你自己配置的地址：模型服务、MCP Server、网络搜索、可选的技能/服务接口。
• 对话内容与笔记只发给你配置的模型服务商。配本地模型 + 本地向量库，可以完全断网运行。

以下披露是为了不让你意外：

• 访问 vault 之外的文件 —— 智能体的笔记工具走 Obsidian API，但部分功能（技能安装、导入导出、内置 HTTP API server）会用 Node 的 fs，可能读写库外路径；不主动触发就不会碰。
• 执行命令 —— 部分技能功能会通过 child_process 跑技能自带的脚本。工具调用有权限确认，但安装来路不明的技能等于运行它的代码。
• 剪贴板 —— 用于聊天里的复制/粘贴操作。
• 技能里的动态代码 —— 安装的技能可能以脚本形式执行，只装你信任的。

已知不足

• DeepSeek 提示缓存的命中率没优化好：prompt 前缀不稳定，缓存收益远低于预期。
• Skills 与 MCP 是雏形：MCP 只实现 HTTP/SSE，没有 stdio，也没在大量真实服务上验证过。
• 大文件技术债：main.ts 与 src/ui/views/view.ts 各几千行，拆分没做完。
• 没有自动化测试：改代码靠人肉验证。
• 内置 HTTP API server 无鉴权：默认关闭、只适合本机，别暴露到公网。
• 移动端能用但慢：大仓库的索引与检索体验一般。

授权

MIT。源码公开，Release 里放的是编译产物。

意见与 bug 走 GitHub Issues。个人项目，不承诺维护节奏。