// the find
NVIDIA-AI-Blueprints/rag
This NVIDIA RAG blueprint serves as a reference solution for a foundational Retrieval Augmented Generation (RAG) pipeline.
NVIDIA's reference RAG implementation built on their NIM microservices stack — Nemotron models for embedding, reranking, and generation, Elasticsearch or Milvus for vector storage, LangChain for orchestration. It's aimed at enterprise teams who want a validated starting point for document Q&A that runs on NVIDIA hardware and doesn't require building the retrieval pipeline from scratch.
The multimodal ingestion pipeline is genuinely useful — it handles tables, charts, infographics, and audio through specialized NIMs rather than dumping everything into a generic text chunker. The agentic RAG mode with LangGraph plan-and-execute is a real architectural choice, not just a marketing label: parallel sub-tasks, scope discovery, and streaming stage events are all wired up. Deployment options are well-covered — Docker Compose for local, Helm for Kubernetes, OpenShift support with an actual values file — so you're not left guessing how to run this in your environment. RAGAS evaluation scripts are included, which means you can actually measure whether your retrieval is working rather than eyeballing outputs.
The NIM dependency is the elephant in the room — you need an NVIDIA NGC API key and GPU hardware that can run these microservices, which rules out most non-enterprise setups and makes local development on a MacBook or CPU-only machine essentially impossible. LangChain as the orchestration layer is a double-edged choice: it's familiar but it abstracts away enough that debugging retrieval failures means wading through chain internals rather than your own code. The Elasticsearch default for vector storage is an odd pick given cuVS/Milvus are available and GPU-accelerated — Elasticsearch works but it's not what you'd choose if GPU throughput is the selling point. The licensing situation is a mess: Apache 2.0 for the code, NVIDIA Community Model License for the NIMs, three different Llama licenses for the Llama-based models, and a separate NVIDIA Asset License for the sample data — legal review before production deployment is not optional.