NVIDIA Blueprints & the Retriever Stack

Accessed/verified: 2026-07-14. This shelf is the how-to/catalog layer for NVIDIA's retrieval and agentic stack; the architecture argument (why each pipeline box exists, the full stage-by-stage NVIDIA mapping table) lives in the Enterprise RAG in Production brief §4 — read that first, then use these pages for the concrete repos, model identifiers, and deploy steps.

The three product concepts

• NIM microservices — "NVIDIA NIM provides prebuilt, optimized inference microservices for rapidly deploying the latest AI models on any NVIDIA-accelerated infrastructure — cloud, data center, workstation, and edge" [nim-page]. Each NIM is an enterprise-grade container packaging a model, an optimized inference engine (TensorRT-LLM, vLLM, or SGLang), industry-standard APIs, and runtime dependencies [nim-page]. One NIM = one dockerizable model endpoint with an OpenAI-compatible API.
• Blueprints — "comprehensive reference workflows featuring NVIDIA acceleration libraries, SDKs, and NIM microservices" [nim-page]; concretely, the github.com/NVIDIA-AI-Blueprints org (37 repositories on the access date [blueprints-org]): each repo is a deployable reference application that wires several NIMs together with compose files, Helm charts, and a UI.
• Launchables — NVIDIA Brev's "one-click deployable environments" that "bundle hardware, software, and code into shareable links" on Brev, NVIDIA's "GPU-accelerated AI/ML development platform" spanning multiple cloud providers [brev-docs]. Blueprints ship launchables as the try-it-on-rented-GPUs path — e.g. the VSS blueprint's Brev launchable runs on 2x RTX PRO 6000 SE [vss-github].

How the pieces relate

flowchart LR
  subgraph retrieval[Retrieval leg — NeMo Retriever]
    I[nv-ingest extraction NIMs] --> E[embedding NIM]
    E --> V[(Elasticsearch / Milvus cuVS)]
    V --> R[reranking NIM]
  end
  R --> G[Generation: Nemotron 3 Super NIM]
  T[NeMo Agent Toolkit: glue, profiling, MCP] -.orchestrates.-> retrieval
  T -.orchestrates.-> G
  S[NeMo Guardrails: input / dialog / retrieval / execution / output rails] -.wraps.-> G

• Retrieval leg — NeMo Retriever: "an end-to-end, agent-ready stack" that "transforms enterprise documents into a structured knowledge layer" [nemo-retriever-page], made of the NeMo Retriever Library (GPU-accelerated ingestion, a.k.a. nv-ingest), Nemotron Retriever open models, and the embedding/reranking/extraction NIMs (extraction, NIM family).
• Generation — Nemotron: the RAG blueprint's default generator is nemotron-3-super-120b-a12b [rag-github]; the model family background (hybrid Mamba-Transformer MoE, Nano/Super/Ultra lineup) is covered in the enterprise-RAG brief §4.4.
• Glue — NeMo Agent Toolkit (NAT): "an open-source library for efficiently connecting and optimizing teams of AI agents" (v1.8), framework-agnostic across LangChain, LlamaIndex, CrewAI, Microsoft Semantic Kernel, Google ADK, and plain Python; ships as pip install nvidia-nat with a nat run / nat serve / nat evaluate CLI, token-level workflow profiling, and MCP support in both directions (client and server) [nat-github]. AI-Q is built on it.
• Safety — NeMo Guardrails: "an open-source toolkit for easily adding programmable guardrails to LLM-based conversational systems," with five rail types — input, dialog, retrieval ("applied to the retrieved chunks in the case of a RAG scenario"), execution, and output [guardrails-github]. The RAG blueprint exposes it as an option; the security argument for retrieval rails is in the brief §5.5.

Verified catalog (this shelf's scope)

Other org members verified to exist but out of this shelf's scope (one-liners from the org listing [blueprints-org]): ai-virtual-assistant (customer-service assistant), pdf-to-podcast, llm-router (route requests to the best model), vulnerability-analysis (container CVE triage), data-flywheel, biomedical-aiq-research-agent (AI-Q derivative), nemotron-voice-agent.

Provenance note: the NeMo Retriever learning-library listing on developer.nvidia.com renders its list client-side (the section headers arrive without items over plain HTTP), so this catalog was assembled from the GitHub org, docs.nvidia.com, and the readable parts of the developer page — every id above comes from one of the cited pages as fetched on the access date.

Sources

nim-page: {title: "NVIDIA NIM Microservices — product page (definition; TensorRT-LLM/vLLM/SGLang engines; blueprints relation)", url: "https://www.nvidia.com/en-us/ai-data-science/products/nim-microservices/", accessed: "2026-07-14"}
blueprints-org: {title: "NVIDIA-AI-Blueprints GitHub organization — repository listing (37 repos)", url: "https://github.com/orgs/NVIDIA-AI-Blueprints/repositories", accessed: "2026-07-14"}
brev-docs: {title: "NVIDIA Brev Documentation — platform and Launchables definitions", url: "https://docs.nvidia.com/brev/latest/", accessed: "2026-07-14"}
nemo-retriever-page: {title: "NVIDIA NeMo Retriever — developer page (agent-ready stack; Library + open models + NIMs)", url: "https://developer.nvidia.com/nemo-retriever", accessed: "2026-07-14"}
rag-github: {title: "NVIDIA RAG Blueprint — NVIDIA-AI-Blueprints/rag (v2.6.0)", url: "https://github.com/NVIDIA-AI-Blueprints/rag", accessed: "2026-07-14"}
aiq-github: {title: "AI-Q NVIDIA Blueprint — NVIDIA-AI-Blueprints/aiq (v2.1.0)", url: "https://github.com/NVIDIA-AI-Blueprints/aiq", accessed: "2026-07-14"}
nv-ingest-github: {title: "NeMo Retriever Library (nv-ingest) — NVIDIA/nv-ingest (26.5.0)", url: "https://github.com/NVIDIA/nv-ingest", accessed: "2026-07-14"}
embed-matrix: {title: "NeMo Retriever Text Embedding NIM — Support Matrix", url: "https://docs.nvidia.com/nim/nemo-retriever/text-embedding/latest/support-matrix.html", accessed: "2026-07-14"}
rerank-matrix: {title: "NeMo Retriever Text Reranking NIM — Support Matrix", url: "https://docs.nvidia.com/nim/nemo-retriever/text-reranking/latest/support-matrix.html", accessed: "2026-07-14"}
vss-github: {title: "Video Search and Summarization Blueprint — NVIDIA-AI-Blueprints/video-search-and-summarization (3.2.0)", url: "https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization", accessed: "2026-07-14"}
nat-github: {title: "NVIDIA NeMo Agent Toolkit — NVIDIA/NeMo-Agent-Toolkit (v1.8, nvidia-nat)", url: "https://github.com/NVIDIA/NeMo-Agent-Toolkit", accessed: "2026-07-14"}
guardrails-github: {title: "NeMo Guardrails — NVIDIA/NeMo-Guardrails (five rail types)", url: "https://github.com/NVIDIA/NeMo-Guardrails", accessed: "2026-07-14"}