Getting Started

Getting Started / 快速开始

English · 中文

English

1. Download the correct package

Open the latest release.

• Choose the asset containing WIN-EXE for normal desktop use.
• Choose the MCP ZIP only when configuring a client that supports local stdio.
• Choose .mcpb for Registry/MCPB-aware clients.

Extract the Windows desktop archive to a normal writable folder and launch OmniClipRAG.exe. Keep the extracted package intact; do not run only the EXE after separating it from its accompanying files.

2. Let startup detection finish

The main window appears before heavier checks finish. Runtime detection starts after the first paint and the Query Console remains unavailable until the result is known. A short detecting state is expected.

If the result says that Runtime or models are missing, continue to the Runtime page instead of repeatedly restarting the application.

3. Choose the data root

The data root stores configuration, logs, indexes, model caches, Runtime payloads, and workspace state. It is separate from both the application folder and your source notes.

Choose a stable local location with enough free disk space. Avoid temporary folders, cloud placeholders that are not always online, and removable drives that may disappear during a build.

4. Prepare Runtime and models

Use the Runtime page to install or repair the CPU/CUDA components and local models. The packaged app includes the installer-side Python needed for this process; normal users do not need a separate system Python installation.

See Runtime and Models for CPU/CUDA selection and repair guidance.

5. Add a Markdown knowledge base

Add the root directory that contains your Markdown notes. Confirm the displayed scope before building, especially when the selected directory contains backups, generated history, or unrelated archives.

6. Run the first full build

The first build parses source files, writes the lexical database, and—when semantic Runtime and models are ready—creates vector data. Large knowledge bases can take time. Use the visible stage and progress text to distinguish active work from a failure.

7. Search and optionally enable live watch

After the index is ready, open the Query Console and run a short natural-language query. Results show their source family and source identity.

Enable live watch only for directories that should follow later edits. Live watch is event-driven and uses a quiet period so repeated saves are merged into one incremental update.

Recommended first verification

1. Search for an exact phrase that exists in one note.
2. Search for a concept described with different wording.
3. Edit one watched Markdown file and wait for the quiet period.
4. Confirm the activity log reports one incremental update.
5. Search for the new content.

中文

1. 下载正确的发行包

进入最新版本。

• 普通桌面用户下载文件名带有 WIN-EXE 的压缩包。
• 只有需要手动配置本地 stdio MCP 客户端时才下载 MCP ZIP。
• 支持 Registry/MCPB 的客户端可以选择 .mcpb。

把 Windows 桌面版完整解压到普通可写目录，然后运行 OmniClipRAG.exe。不要只把 EXE 单独拿出来运行，它需要同目录中的配套文件。

2. 等待启动检测完成

主窗口会先显示，较重的检查随后在后台进行。Runtime 检测完成前查询台会保持不可用，短暂显示“检测中”属于正常现象。

如果结果提示缺少 Runtime 或模型，请进入 Runtime 页面处理，不需要反复重启程序。

3. 选择数据目录

数据目录用于保存配置、日志、索引、模型缓存、Runtime 和工作区状态。它既不是程序目录，也不是原始笔记目录。

请选择空间充足、长期稳定的本地位置。不要使用临时目录、未始终保存在本机的云占位目录，或者建库过程中可能断开的移动盘。

4. 准备 Runtime 和模型

在 Runtime 页面安装或修复 CPU/CUDA 组件与本地模型。发布包已经带有安装过程所需的 Python，普通用户不需要额外安装系统 Python。

CPU/CUDA 选择和修复方法见：Runtime 与模型。

5. 添加 Markdown 知识库

添加包含 Markdown 笔记的根目录。建库前确认界面显示的范围，尤其注意所选目录中是否混有备份、自动生成历史或无关归档。

6. 执行第一次全量建库

第一次建库会解析来源文件、建立字面数据库，并在语义 Runtime 与模型就绪时建立向量索引。大型知识库需要一定时间，应通过界面的阶段和进度文字判断任务仍在运行还是已经失败。

7. 查询并按需开启热监听

索引就绪后，在查询台输入简短自然语言问题。结果会显示来源类型和来源文件。

只有确实需要持续跟踪修改的目录才开启热监听。热监听由真实文件事件触发，并等待安静时间，把连续保存合并为一次增量更新。

建议完成的第一次验证

1. 搜索笔记中确实存在的一句原文。
2. 换一种说法搜索相同概念。
3. 修改一个已经监听的 Markdown 文件并等待安静时间。
4. 确认活动日志只报告一次增量更新。
5. 搜索刚添加的新内容。