AutoRAG is an open-source framework from Marker Inc in Korea for automatically evaluating and optimizing retrieval-augmented generation pipelines. Rather than guessing which chunking strategy, retriever or embedding model works best for your documents, AutoRAG tests the combinations systematically and finds the pipeline that performs best on your data. It is actively developed on GitHub, with more than 5,100 stars.
Building a good RAG system involves dozens of interacting choices: how documents are parsed and chunked, which embedding model and vector store to use, whether to rerank or rewrite queries, and which generator settings to apply. AutoRAG structures this search. It starts from your raw documents, generates QA evaluation data, then runs modular pipeline experiments across parsing, chunking, retrieval, prompt making and generation modules, scoring each configuration with retrieval and generation metrics. The result is a reproducible, evidence-backed RAG pipeline instead of trial-and-error tuning, exportable for deployment in your own application.
Experiments are configured through readable YAML files, and results can be inspected side by side to see exactly which pipeline stage helped or hurt. AutoRAG works with Korean and English corpora and popular vector databases, and its parsing stage preserves document structure such as tables, which is where many naive RAG pipelines quietly lose accuracy.
AutoRAG is free and open source under the Apache-2.0 license. You run it on your own machine and pay only for any LLM or embedding API calls your experiments make.
AutoRAG is for developers and data teams building document question-answering and enterprise search products who want to choose RAG components based on measured performance rather than folklore.