// the take
UltraRAG is an MCP-based framework for assembling RAG pipelines out of swappable components (retriever, generation, reranker, evaluation, memory), each running as its own MCP server and wired together via YAML. It's aimed at RAG researchers who want to reproduce papers and run benchmarks without rewriting pipeline glue, plus a UI layer for building and demoing pipelines without touching code at all.
Splitting retriever/generation/reranker/evaluation into independent MCP servers means you can swap a backend (Faiss, Milvus, Qdrant for retrieval; vLLM or any LiteLLM-backed model for generation) without touching pipeline logic elsewhere. The YAML orchestration supports loops and conditionals, so the example pipelines actually implement real iterative RAG methods (IRCoT, Search-R1, RankCoT) as runnable configs rather than just linking to papers. The evaluation server standardizes benchmark metrics across pipelines, which is the kind of thing that normally gets rewritten ad hoc in every RAG research repo. The UI's pipeline builder syncs both ways with the underlying YAML/code, so it's not just a toy chat frontend bolted on top.
Full install drags in vLLM, transformers, and multiple vector DB clients at once, and the GPU build is pinned to CUDA 12.9 specifically — expect real dependency wrangling if your stack doesn't match exactly. Running each component as a separate MCP server is more processes to deploy, monitor, and debug than a single-process RAG script, and MCP's production tooling (logging, failure isolation across servers) is still immature industry-wide, not just here. The project has gone through two major rewrites in under a year (v1 to v2 to v3), with old versions kept alive only as separate branches — that's a lot of API churn to track if you're building on top of it rather than just running the included pipelines. The UI ships its own auth and chat-storage layer (ui/backend/auth.py, chat_store.py), which is new code worth auditing carefully before exposing it beyond a local research box, despite docs describing a 'production environment deployment' path.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →