NVIDIA RAG Blueprint
A reference retrieval-augmented generation pipeline built on NVIDIA NIM microservices, with GPU-accelerated ingestion and search, a reference UI, and an optional agentic mode for multi-hop questions.
NVIDIA RAG Blueprint is a reference solution and starting point for building retrieval-augmented generation pipelines with NVIDIA NIM microservices. It lets an organization answer natural-language questions grounded in its own data, combining GPU-accelerated components with NeMo Retriever models, multimodal and vision-language models, and guardrailing services in a design meant to be decomposed and configured for enterprise governance, latency and scale needs.
Data ingestion handles documents containing text, tables, charts, infographics and audio, with custom metadata, and the search layer supports multi-collection search, hybrid dense and sparse retrieval, reranking and a pluggable vector database. For complex, multi-hop or ambiguous questions, an optional Agentic RAG mode adds a LangGraph plan-and-execute pipeline alongside the standard retrieve-then-generate chain, with scope discovery, parallel sub-tasks, synthesis and optional verification, streamed as stage events in the reference UI and API.
It ships with open-source code and a pre-built reference UI, and can be deployed with local Docker or on Kubernetes.