AutoRAG
Maintained by Marker Inc. Korea, MIT licensed.
The project pivoted
AutoRAG was widely catalogued as “an AutoML tool for RAG — automatically optimize your RAG pipeline with a single YAML file”. AutoRAG 2.0 is a different product: a self-evolving librarian agent for document collections, shipped as
@autorag/librarianon npm. Older descriptions of it as pipeline AutoML no longer describe what the repository contains.
🔗 https://github.com/Marker-Inc-Korea/AutoRAG
What It Is Now
Rather than returning raw file paths and line numbers, it acts as a librarian that retrieves, reads, evaluates and curates findings into structured knowledge units — searching across PDFs, wikis, notes and papers, then synthesizing numbered findings with locations. It learns from user feedback which retrieval methods work for which query types.
| Module | Role |
|---|---|
| MinSync | Local indexing layer for BM25 / vector / hybrid retrieval |
| Jikji | Discovery and initial ranking |
| Pi Agent Loop | The retrieve-and-curate reasoning cycle |
| Datasource Skills | Connectors — Slack, Discord, Gmail, GitHub, cloud drives, Obsidian |
| dupey | Duplicate document detection |
Runtime requires Node ≥24 or Bun; PDF parsing needs Java 11+.
Why It Matters Here
The pivot is itself the interesting datum: a project that set out to auto-tune retrieval hyperparameters ended up as an agent that decides retrieval strategy at query time. That is the same movement described in Agentic Search — from a tuned static pipeline to a policy chosen per query.