Personal RAG
← Back to Personal RAG Feedback
/

MEMORY FOR YOUR AGENT · README

Personal RAG

v0.1.0

What it does, how to install it, how to use it, and where your data lives — the whole manual on one page.

它是什么、怎么装、怎么用、数据在哪——整本手册就在这一页。

Why Personal RAG为什么是 Personal RAG

Your agent can search the open web. It cannot read your own notes, meetings and course transcripts — unless something indexes them.

Personal RAG does exactly that and nothing else: a folder of Markdown in, cited answers out, on your machine. The knowledge base is a folder, not an account.

你的 agent 能搜整个公网，但读不到你自己的笔记、会议与课程转写——除非有人把它们索引起来。

Personal RAG 只做这件事：一个 Markdown 文件夹进去，带出处的答案出来，全程在本机。知识库是一个文件夹，不是一个账号。

What it does能做什么

  • Point it at any folder of Markdown; no Obsidian required, compatible if you use it.
  • Ingest and index, manually or automatically as files appear.
  • Search or ask; every answer carries its source file and passage.
  • Four front doors: desktop UI, CLI, local HTTP API, and an MCP server for agents.
  • Maintenance tools: reindex, audit, doctor, and query labelling for evaluation.
  • 指向任意 Markdown 文件夹；不用 Obsidian，用也完全兼容。
  • 入库与建索引，手动或自动盯着文件夹。
  • 搜索或提问，每个答案都带出处：哪个文件、哪一段。
  • 四个前门：桌面 UI、CLI、本地 HTTP API、给 agent 的 MCP server。
  • 维护工具：reindex、audit、doctor，以及给评测用的查询标注。

Install安装

Windows installer. After launching, pick the vault folder (your notes) and a data folder for the index — keep the data folder outside the vault.

Windows 安装包。启动后选 vault 目录（你的笔记）和索引数据目录——数据目录不要放进 vault 里。

How to use使用说明

  1. Settings: choose vault and data folder, save.
  2. Turn on auto-ingest, or press ingest then index.
  3. Ask a question in the search box; check the citation under each hit.
  4. For an agent: configure your MCP client to run personal-rag mcp.
  5. Scripts can use the CLI or the local API.
  1. 设置里选 vault 与数据目录，保存。
  2. 打开自动入库，或手动点 ingest 再 index。
  3. 在搜索框提问，看每条结果的出处。
  4. 给 agent 用：MCP 客户端里配 personal-rag mcp。
  5. 脚本可以用 CLI 或本地 API。

Data, privacy, limits数据、隐私、限制

The vault is the source of truth; the index is derived and can be rebuilt. Do not delete the data folder by hand — run reindex instead.

Retrieval and indexing run locally. Changing chunking or embedding parameters requires bumping the index version and rebuilding, otherwise results silently degrade.

vault 是真相源，索引是派生的，可以重建。别手动删数据目录——用 reindex。

检索与索引都在本地。改切分或嵌入参数必须提升索引版本并重建，否则结果会悄悄变差。

Version, update, tech版本、更新、实现

Version 0.1.0. Updates: the desktop app checks a manifest at pra-updates.bimon.net and installs only after you confirm.

Rust; SQLite plus a vector store and a keyword index; MCP server over stdio; evaluation sets for both human-style and agent-style queries.

版本 0.1.0。更新：桌面端读 pra-updates.bimon.net 的清单，你确认后才装。

技术：Rust；SQLite + 向量库 + 关键词索引；MCP server 走 stdio；自带人话与 agent 风格两套评测集。