Marker-Inc-Korea Releases 'AutoRAG 2.0', a Local Data Unified Search Agent

Marker-Inc-Korea has moved its existing Python-based RAG AutoML project (1.x) to legacy and released AutoRAG 2.0, which searches all information on a computer (files, messenger chat logs, emails, Notion, etc.).

Key Features and Architecture

• In-place Federation: Connects directly to each tool's native store (katok, discrawl, mailcrawl, etc.) without data migration, maintaining access control over original paths and privacy.
• Single Model Loop: Instead of multi-agent layers, a single model owns the retrieval-read-curation loop to generate low-latency responses in BM25, vector, and hybrid modes.
• Self-evolving: Performance improves with usage frequency, prioritizing past successful methods (check_memory) and learning from feedback.
• Zero-Config: Built-in MinSync and local embedding eliminate the need for external API keys or vector DB tuning.

Installation and Tech Stack

• Installation: npm install @autorag/agent (Requires Node >= 24.0.0, Bun >= 1.2)
• License: AutoRAG 2.0 is MIT, legacy Python version is Apache 2.0
• Tech Stack: Uses TypeScript, Bun, Vitest, Biome
• Interface: Provides CLI, Web Dashboard, and skills directory for AI Agents