AutoRAG
Self-evolving librarian agent for document collections and local RAG retrieval.
Federates local stores, reads source files directly, and curates structured answers.
AutoRAG 2.0 is a specialized search agent built on a customized Pi agent loop that retrieves candidates, reads source files directly, judges evidence, and curates answers. It suits developers and professionals managing local document collections who need direct answers instead of raw grep dumps.
What it is
It searches PDFs, wikis, notes, and knowledge bases using pluggable retrieval methods like BM25, vector, and hybrid search. A thin PDF extraction retry uses OpenDataLoader's docling-fast backend for sparse markdown, while Jikji provides find-first discovery and dupey detects document families.
Who it's for
It suits data scientists and engineers looking for an autonomous librarian to search and synthesize local documentation without central index migration. It does not suit teams wanting a multi-model coordinator or a SaaS-managed vector database setup.
Key features
In-Place Federation
Federates CLI-owned stores in place without migrating data to a central index.
Self-Evolving Memory
Learns productive document areas and successful retrieval strategies over time from usage.
Pluggable Retrieval
Supports grep, BM25, vector search, and hybrid modes over a shared MinSync chunk lifecycle.
Thin PDF Extraction
Automatically retries sparse multi-page PDFs using the docling-fast hybrid backend.
Up and running in 3 steps
Our honest verdict
What's great
- Federates local stores in place without forcing data migration to a central index
- Self-evolving memory system learns effective retrieval strategies from actual usage
- Pluggable local retrieval methods handle plain text, research papers, and legal documents
- Reads source files directly through its built-in bash tool to verify evidence
Where it falls short
- Requires configuring an authenticated runtime model provider since it does not ship private defaults
- MinSync and external binaries like Jikji and dupey require local setup and maintenance
- Thin PDF extraction relies on the availability of the docling-fast hybrid sidecar
Pricing & license
The only cost is running it yourself, as it is an open-source tool utilizing your own authenticated model provider runtime.
Open remote AI-training roles
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
What happened to the original Python-based AutoRAG?
The original RAG AutoML pipeline optimization tool now lives in the legacy/ directory and remains in maintenance mode, while new development focuses on AutoRAG 2.0.
Does AutoRAG upload my documents to a third-party server?
No. AutoRAG federates stores in place and runs retrieval locally over MinSync chunks without a central index server holding your corpus.
How does AutoRAG handle sparse or unreadable PDFs?
Its default PDF parser checks multi-page quality and retries sparse markdown through OpenDataLoader's docling-fast hybrid backend.
