Stop Reading File Paths: AutoRAG Turns Your Document Search Into a Librarian

You know the drill. You search your company's wiki, your PDFs, or your research papers, and you get back a wall of file paths and line numbers. Then the real work begins: opening each file, reading the context, synthesizing the answer yourself. It's exhausting, and it's exactly the problem AutoRAG wants to eliminate.

AutoRAG is a self-evolving librarian agent for document collections. Instead of dumping raw search results on you, it searches, reads, thinks, and comes back with clean, numbered knowledge units. It's built as a customized Pi agent, configured specifically for this librarian role. The project lives at Marker-Inc-Korea/AutoRAG, and it's a complete reimagining of the original AutoRAG tool.

What It Does

AutoRAG is not a search tool in the traditional sense. You ask it a question like "What were the key findings in the Q3 report?" and instead of giving you a list of matching lines, it gives you something like this:

[1] Revenue grew 23% YoY to $4.2M, driven by enterprise contracts. (pages 3-5)
[2] Three new risk factors: supply chain, regulatory, talent retention. (pages 12-14)
[3] Headcount target missed by 12 — engineering hiring bottleneck. (page 8)

No file paths. No line numbers. Just curated, actionable knowledge with page references.

Under the hood, it uses a two-tier workflow. A parent orchestrator agent delegates exploration to explorer agents. The orchestrator owns process-bound retrieval tools and hands out bounded seed packs to the explorers, who then use read-only tools like read, grep, find, and ls to inspect the actual documents. The roles and providers are independently configured from the models available in your authenticated runtime—AutoRAG doesn't ship with a private provider default.

The retrieval side supports pluggable methods. It ships with two enabled by default: lexical BM25 and semantic MinSync, both wired through a RetrievalMethodRegistry. A ResultMerger handles cross-method score normalization and deduplication, so you get one unified result set regardless of how many methods contributed. The architecture is ready for additional vector and hybrid backends.

Why It's Cool

It shifts the cognitive load from you to the agent. Every search tool makes you do the synthesis. AutoRAG does it for you. That's a fundamental shift in how you interact with your documents—you're not just querying, you're delegating.

It actually gets smarter over time. This is the standout feature. AutoRAG has a self-evolving memory system. Every search teaches it something: which retrieval methods work for which query types, which document areas are most productive, and what you found useful through explicit feedback. A fresh AutoRAG tries everything; a seasoned one knows exactly where to look. This isn't a static configuration—it's learned behavior from real usage.

It respects that different documents need different search strategies. The README lays this out nicely:

• Plain text and config files? Grep works fine—fast, precise, literal.
• Research papers and dense prose? Vector search understands meaning, not just keywords.
• Legal documents and specs? BM25 handles domain terminology well.
• Mixed collections? Hybrid approaches combine precision and recall.

AutoRAG supports all of these through its pluggable retrieval registry. You don't have to pick one strategy and hope for the best.

It's honest about its lineage. The README explicitly notes that the original AutoRAG (the RAG AutoML pipeline optimization tool) still lives in the legacy/ directory. It's not abandoned—it continues to get bug fixes and PyPI releases. But new feature development is focused on this 2.0 version. That's a refreshingly clear migration story for existing users.

How to Try It

Head over to the repository at Marker-Inc-Korea/AutoRAG to get started.

The key thing to know is that BM25 and MinSync retrieval methods are enabled by default—you don't need explicit configuration for standard lexical and semantic retrieval. If you want to disable them, you can set "bm25": false (and presumably the equivalent for MinSync) in your configuration.

Since AutoRAG is a customized Pi agent, you'll want to check the README for the specifics on setting up your runtime and configuring your model providers. The roles and providers are configured independently from the models available in your authenticated environment, so you'll need to have those set up first.

If you're coming from the original AutoRAG, don't worry—the legacy version still works exactly as before. Just check the legacy README for its documentation and keep filing issues in the same repository.

Final Thoughts

AutoRAG is for anyone who's tired of being the human search engine—the person who gets a grep dump and has to turn it into actual understanding. It's particularly useful for teams with large document collections: PDFs, wikis, research papers, knowledge bases. The self-evolving memory is the kind of feature that sounds like a gimmick until you realize how much time it saves when the agent learns where your most productive information lives.

It's early days for a 2.0 rewrite, so expect some rough edges, but the core idea is solid. Stop reading file paths. Start getting answers.

Follow @githubprojects for more developer tools and open source projects.