UltraRAG is a low-code framework from OpenBMB for building complex retrieval-augmented generation pipelines. Built around the Model Context Protocol, it lets developers compose retrievers, rerankers, generators and evaluators into full RAG systems through simple configuration instead of glue code. The project has more than 5,700 stars on GitHub and was actively developed in 2026.
Serious RAG systems go far beyond a single vector lookup: they rewrite queries, retrieve in multiple hops, reason over intermediate results and check their own outputs. UltraRAG provides modular components for each of those stages, including embedding, retrieval and generation, which snap together into pipelines you can inspect and extend. It supports multimodal retrieval with vision-language models, works with popular LLMs such as Qwen and DeepSeek, integrates serving engines like vLLM, and ships with a user interface and evaluation tooling. The framework comes from OpenBMB, the group behind a series of widely used open model projects, and reflects their focus on techniques that can be reproduced outside large industrial labs. Example pipelines in the repository demonstrate patterns such as query decomposition and iterative retrieval that are otherwise described only in papers, making UltraRAG a hands-on way to learn and benchmark modern RAG design. Its component model also makes ablation studies and fair framework comparisons straightforward to run.
Free. UltraRAG is open source under the Apache-2.0 license; running it only requires your own infrastructure and model access.
Developers and researchers building advanced retrieval systems, such as deep-research assistants and enterprise knowledge bases, who have outgrown basic RAG tutorials and frameworks.