AutoRAG

AutoRAG, from Marker-Inc-Korea, is an open-source librarian agent that searches a user's files and connected datasources and answers questions with cited sources. The repository now hosts AutoRAG 2.0 (AutoRAG Agent), published as the @autorag/librarian package. The original Python RAG AutoML tool lives in the legacy/ directory and is described as maintained in maintenance mode.

What It Is

AutoRAG Agent is a retrieval agent built on the Pi agent framework. A single configured model searches across several retrieval methods, opens source documents directly with bash to verify facts, and returns a plain answer whose inline citations are resolved into numbered, source-backed results. It is available as a CLI, a terminal UI, a TypeScript API, an MCP server, and agent skills. A desktop app, whose source lives in a separate repository, wraps the same engine in a file manager.

How Retrieval Works

The README describes federating data in place instead of copying it into a central vector database. Retrieval methods include BM25 lexical search, semantic vector search, hybrid search with Reciprocal Rank Fusion, Jikji find-first discovery, and file-name search (Everything on Windows, FSearch on macOS/Linux). Results are merged and deduplicated, then the agent reads promising files directly. Cited evidence judged to support an answer is stored in a local memory file and offered as advisory context for later searches.

Datasources and Document Parsing

Connected sources include local documents, KakaoTalk, Discord, WhatsApp, Telegram, Slack, Notion, email archives, Obsidian vaults, GitHub issues and pull requests, GitHub Gists, cloud drives via rclone, photos and screenshots, RSS, and macOS Spotlight. Documents such as HWP/HWPX, PDF, DOCX, XLSX, PPTX, EML and plain text are parsed in-process without Java. OCR for images and scanned pages is optional and off by default.

Local-First and Model Options

Embeddings run locally with the qwen3-embedding-0.6b model, and the README states file contents are not sent to a remote embedder. The librarian needs one chat model, which can come from a user's own provider or a hosted AutoRAG model provider. AutoRAG Lite offers the same indexing and retrieval engine without an LLM, exposed through an MCP server for tools such as Claude Code and Codex.

Setup Path

Install with bun install -g @autorag/librarian or npm (Node.js 24 or later, or Bun). Then run autorag init --search-paths <folder>, autorag refresh, and autorag search "<question>", or launch autorag tui to sign in and choose a model.