Marker-Inc-Korea/AutoRAG

Overview

AutoRAG Agent is a personal librarian agent for office work: it reads files, mail, chats and notes in place, searches them with a mix of sparse, vector and hybrid retrieval, and returns answers where each claim cites the source passage AutoRAG Agent. It is built on the Pi agent framework and can also act on files when asked, such as organizing, renaming, moving or editing them while recording each change.

The engine ships as the npm package @autorag/librarian; the separate desktop app (Dazzi) is an Electron + React client built on the same engine packaging. External datasources such as KakaoTalk, Slack, Discord, Telegram, WhatsApp, mail and GitHub are queried through their own CLI tools or archives rather than copied into a central vector database supported datasources.

AutoRAG Lite is a model-free mode of the same engine: local sparse (SPLADE) and vector search with no LLM tokens, exposed through an autorag-mcp stdio server that other agents can call AutoRAG Lite.

Quick Start

Requires Node.js ≥ 24 or Bun, and a chat model for the full agent. Install globally:

bun install -g @autorag/librarian
# or: npm install -g @autorag/librarian

Initialize a workspace and build the index, then check health initialize and index:

autorag init --search-paths ~/Documents/research
autorag refresh
autorag status

autorag init does not choose a model, so select one on first run before searching. Open the TUI, sign in and pick a model, then ask a question search from CLI:

autorag tui      # then /login and /model
autorag health
autorag search "What are our primary Q3 deliverables?"

For model-free retrieval, initialize a Lite workspace instead and run the MCP server:

autorag lite init --search-paths ./documents
autorag lite refresh
AUTORAG_CONFIG=/absolute/path/to/.autorag/config.json autorag-mcp

Configuration and Constraints

• Languages and parsers. One global languages setting drives parsing, defaulting to ko,en; override with --languages, AUTORAG_LANGUAGES or the config file. HWP/HWPX, PDF, DOCX and image OCR go through kordoc; spreadsheets, CSV, PPTX and EML have built-in readers parsers.
• OCR is opt-in. parserOptions.ocr.enabled defaults to false; image files are not parsed and scanned pages stay unrecognized until enabled. The built-in PP-OCRv5 model covers Korean and English only; other languages use Tesseract, downloaded per language OCR.
• Parse isolation. Parses run in a worker thread with fileTimeoutMs (default 120000) and optional maxHeapMb; a timed-out, out-of-memory or crashed parse skips one file and the rest of the corpus still indexes parse bounds.
• Auto-downloaded tools. MinSync is installed into <workspace>/.autorag/bin on first run; filee-cli (~85 MB, SHA-256 verified) is fetched for format conversion; Windows bundles a private Everything instance and FSearch must be installed separately on macOS/Linux prerequisites.
• Exclusions are literal paths, not globs. Repeat --exclude for each path; excluding a folder excludes descendants exclusions.
• Model access. You can use the hosted Dazzi Models provider via /login or connect your own provider key and pay the provider directly; there is no API-key environment variable for Dazzi hosted plan.
• Licensing. AutoRAG 2.0 and the legacy engine are Apache-2.0; the desktop app source is Fair Source (FSL-1.1-MIT), and the bundled third-party binaries keep their own licenses.

HotRepo's take

On the TypeScript daily list for 2026-11-07, AutoRAG gained 40 stars and reached position 16, with 5,397 stars and 485 forks overall. That single snapshot is all the site has recorded, so it supports a statement about that day's visibility rather than a trend. What stands out from the material is scope: the same engine ships as a CLI, a TUI, an MCP server and the basis of a separate desktop app, and the Lite mode lets external agents query local files without any model configuration. The layout — data federated in place, retrieval memory stored under ~/.autorag/memory.json, and agent-driven file edits routed to the OS trash with old → new path records — is worth reading before adopting it for sensitive document folders.