UltraRAG is the first lightweight RAG development framework built on the Model Context Protocol (MCP) architecture, jointly created by THUNLP at Tsinghua University, NEUIR at Northeastern University, OpenBMB, and AI9stars. It enables developers to build complex retrieval-augmented generation pipelines with minimal code through YAML configuration. ## MCP-Based Modular Architecture UltraRAG standardizes core RAG components -- Retriever, Generator, Evaluator -- as independent MCP Servers. Each component is registered at the function level, enabling high reusability across different pipeline configurations. The MCP Client orchestrates these servers, allowing developers to construct complex control structures including conditional branches, loops, and sequential flows through simple YAML definitions rather than writing extensive orchestration code. This architecture means switching from one retrieval strategy to another requires changing a few lines of YAML rather than rewriting application logic. The modular approach also makes it straightforward to test individual components in isolation. ## Low-Code Pipeline Orchestration The framework's YAML-based pipeline definition is its most practical feature. Complex RAG workflows that would traditionally require hundreds of lines of Python can be expressed in dozens of lines of configuration. UltraRAG supports sequential flows, loops, and conditional branches natively within this configuration language. Version 3.0 introduced a WYSIWYG Pipeline Builder with dual-mode editing: a visual Canvas mode and a Code mode that sync in real-time. Clicking "Build" transforms static pipeline logic into an interactive web interface, eliminating the tedious work of UI development for research prototypes. ## Unified Evaluation Framework UltraRAG includes standardized evaluation workflows and integrates mainstream research benchmarks out of the box. This addresses a common pain point in RAG development where researchers spend significant time building custom evaluation harnesses. The unified evaluation system improves experiment reproducibility and makes cross-method comparison significantly more efficient. ## Rapid Prototyping and Deployment Single-command deployment converts pipeline logic into interactive web interfaces. This feature is particularly valuable for academic researchers who need to demonstrate their RAG innovations without investing weeks in frontend development. Docker deployment is also supported with pre-built CPU and GPU images that automatically launch the UI on port 5050. ## AgentCPM Integration UltraRAG integrates with AgentCPM-Report, a localized deep research model that enables on-premises research capabilities without depending on external API services. This addresses privacy and cost concerns for organizations that need RAG capabilities but cannot send data to third-party services. ## Installation and Setup The recommended installation uses the uv package manager for faster dependency resolution. Source installation supports core, full, or on-demand configurations via uv sync with various extras. The framework runs on Windows, macOS, and Linux, with Docker providing the simplest deployment path for production use.