R2R is an open-source, production-ready AI retrieval system from SciPhi, published on GitHub with around 8,000 stars. It packages everything needed to build serious retrieval-augmented generation (RAG) applications — ingestion, hybrid search, knowledge graphs, and an agentic reasoning layer — behind one RESTful API.
Most RAG demos stop at embedding chunks and a vector search. R2R is built for the production version of that problem. It ingests multimodal content (text, PDF, JSON, images, audio and more), indexes it for hybrid search that fuses semantic similarity with keyword matching via reciprocal rank fusion, and automatically extracts entities and relationships into knowledge graphs that improve retrieval quality. On top sits an agentic layer: a reasoning agent, plus a Deep Research API that performs multi-step research across your knowledge base and the web to answer complex questions with citations. The system ships with user and access management, Python and JavaScript SDKs, and flexible deployment — a light pip install for development or a full Docker deployment with Postgres for production. R2R is MIT licensed and self-hosted, with its latest repository activity in late 2025.
R2R is free and open source under the MIT license. It is self-hosted software; your costs are your own infrastructure plus any LLM and embedding API usage you configure.
R2R is for engineering teams building document Q&A, enterprise search, and research assistants who have outgrown a LangChain demo and need an integrated, self-hostable retrieval platform with graphs and agents built in.