AutoRAG

What it is

AutoRAG is a self-evolving librarian agent for document collections. The repository now hosts AutoRAG 2.0, a reimagined version, while the legacy Python-based AutoRAG (RAG AutoML tool) lives in the legacy/ directory. It searches configured source directories and external datasources, then returns curated results rather than raw file paths. It ships with pluggable retrieval methods (BM25, MinSync) and supports data sources via a datasourceSkills mechanism. The project targets non-code document retrieval but can also index code repositories to some extent through the explorer's grep utility.

How it works

AutoRAG uses a two-tier workflow: a parent orchestrator delegates exploration to explorer agents. It supports multiple retrieval methods (BM25, MinSync) and a retrieval pipeline that can combine lexical and semantic search. Datasource skills enable indexing and querying external sources through a shared retrieval pipeline, with security defaults that default-deny datasource access unless explicitly allowed. It reads from a configured workspace, uses a BM25 lexical index (Korean-aware prefix matching) for queries, and returns curated results with source mappings. It can optionally use Jikji for local discovery and indexing, via a find-first strategy, managed through a policy-aware tool integration. Programmatic usage examples show instantiating AutoRAGAgent with searchPaths and datasourceSkills, and calling prepareJikji/refresh or searchDatasourceDocuments.

Getting started

Install from PyPI for legacy AutoRAG: pip install AutoRAG. Code examples show configuring datasource skills and agent setup. For KakaoTalk datasource: create a KatokSkill and pass it to AutoRAGAgent with appropriate datasourceAccess tags/scopes. Example snippet:

import { AutoRAGAgent, KatokSkill } from "@autorag/librarian";
const kakao = new KatokSkill({
  instanceId: "personal",
  tags: ["kakaotalk", "personal", "pii"],
  // Optional: client: new KatokClient({ binaryPath: "katok" })
});

const agent = new AutoRAGAgent({
  searchPaths: ["/path/to/documents"],
  datasourceSkills: [kakao],
  datasourceAccess: {
    allowedTags: ["kakaotalk"],
    allowedScopes: ["/kakao/personal/**"],
  },
});

await agent.refresh(); // refreshes parsed mirrors, BM25/MinSync, and datasource indexes
const results = await agent.searchDatasourceDocuments("meeting with Mina", { topK: 5 });

Configure datasources in a config.json or via programmatic setup, per the documented examples for Slack, GitHub, Gmail, gDrive, Obsidian, RSS, etc. The README and docs describe how credentials are supplied via environment variables and how access controls are configured.

Recent releases

Latest releases include:

• v2.1.0 AutoRAG v2.1.0 (2026-07-24): changes include support for latest OpenAI models in the generator path and dependency updates.
• legacy-v0.3.24 AutoRAG Legacy legacy-v0.3.24 (2026-07-22): maintenance release for legacy Python AutoRAG (legacy/).
• v2.0.0 AutoRAG v2.0.0 (2026-07-20): merges AutoRAG 2.0 as primary project and moves legacy AutoRAG to legacy/.
• legacy-v0.3.23 AutoRAG Legacy legacy-v0.3.23 (2026-07-20): maintenance release for legacy Python AutoRAG (legacy/).
• v0.3.22 (2026-04-03): feature updates including a new NVIDIA reranker module and MiniMax LLM as a first-class generator module.

Install from PyPI: pip install AutoRAG.

Traction

The repository lists 4966 stars and 419 forks.

Behind the repo

Not included in the provided facts block.

Caveats

License is not listed. Created: 2024-01-10. Last push: 2026-08-01. Languages: TypeScript. Documentation notes legacy path and multiple configuration options, including config.json, datasourceSkills, and environment-based credentials. The two major branches are AutoRAG 2.0 and legacy AutoRAG in legacy/. The README indicates the legacy AutoRAG is maintained in maintenance mode with PyPI releases via pip install AutoRAG and references a legacy README for its documentation.