FlashRAG is a Python toolkit for efficient retrieval-augmented generation research, developed by the NLPIR lab at Renmin University of China. It packages the whole RAG pipeline โ indexing, retrieval, reranking, generation and evaluation โ into modular components with dozens of prebuilt pipelines and benchmark datasets, so researchers can reproduce published methods and test new ideas quickly. The project has around 3,600 stars on GitHub, was presented as a WWW 2025 resource, and is MIT licensed.
RAG research is plagued by incompatible codebases: every paper ships its own retriever wrappers and evaluation scripts. FlashRAG standardises this with a config-driven framework in which retrievers (sparse BM25 and dense models), generators and refiners can be swapped with a few lines of configuration. It ships with a large collection of benchmark datasets for tasks such as open-domain QA, multi-hop reasoning and fact verification, along with standard metrics, so a new method can be compared against established baselines like vanilla RAG, Self-RAG-style pipelines and iterative retrieval under identical conditions. Efficient indexing and caching keep experiments fast even over Wikipedia-scale corpora.
FlashRAG is completely free and open source under the MIT license. Costs are limited to your own compute and any hosted model APIs you choose to evaluate.
NLP researchers and engineers who need a trustworthy testbed for RAG methods โ benchmarking retrievers, comparing generation strategies or prototyping a production retrieval pipeline on solid experimental ground.