NVIDIA GenerativeAIExamples is an Apache-2.0 collection of Jupyter notebook reference workflows for building RAG pipelines, agentic workflows and fine-tuning jobs on NVIDIA accelerated infrastructure and microservice architecture, aimed at developers who are already working inside the NVIDIA software ecosystem.

What it is

The repository is a starting point for developers who want to integrate with the NVIDIA software ecosystem to speed up their generative AI systems. It ships reference workflows, sample code and runnable Jupyter notebooks rather than a library or a framework of its own, and it covers RAG pipelines, agentic workflows and model fine-tuning. The material is organised around NVIDIA components such as NIM microservices, NeMo Microservices (NeMo Datastore, NeMo Entity Store, NeMo Customizer, NeMo Evaluator and NeMo Guardrails), NeMo Retriever, NeMo Auditor, TensorRT-LLM and the Triton Inference Server, with the RAPIDS ecosystem used for large-scale data processing.

The concrete problem it solves is the blank-notebook problem. A developer who wants an accelerated retrieval or agent pipeline today has to assemble the model server, the retriever, the guardrails and the evaluation loop by hand before any application logic gets written. GenerativeAIExamples replaces that assembly work with worked examples: a tool-calling fine-tuning, inference, evaluation and guardrailing loop under nemo/data-flywheel/tool-calling, an embedding fine-tuning path under nemo/data-flywheel/embedding-finetuning/, and a GPU-accelerated knowledge graph pipeline under community/knowledge_graph_rag that stands in for vector-only retrieval over large datasets.

Key capabilities

• Agentic RAG pipelines built on Llama 3.1 and NVIDIA NeMo Retriever NIM microservices, demonstrated in RAG/notebooks/langchain/agentic_rag_with_nemo_retriever_nim.ipynb.
• Data Flywheel tutorials that chain NeMo Customizer, NeMo Evaluator, NeMo Datastore, NeMo Entity Store and NeMo Guardrails microservices with NVIDIA NIMs for fine-tuning and evaluation.
• Safety tooling: auditing large language models for vulnerability to unsafe prompts with NeMo Auditor in nemo/NeMo-Auditor/Getting_Started_With_NeMo_Auditor.ipynb.
• Parallel rails inference in nemo/NeMo-Guardrails/Parallel_Rails_Tutorial.ipynb, run to reduce latency and improve throughput.
• Knowledge graph construction and querying for RAG, using RAPIDS to process large-scale datasets efficiently.
• Event-driven RAG for CVE analysis, integrating NVIDIA Morpheus, NIM microservices and RAG pipelines into LLM-based agent pipelines under experimental/event-driven-rag-cve-analysis.
• Vision NIM workflows, including VLM-based monitoring of a video stream for custom events at nim_workflows/vlm_alerts/README.md and natural-language image search with NV-CLIP.

Who uses it and how

• Application developers who already target NVIDIA hardware and want a working RAG or agent pipeline to modify rather than a framework to learn from scratch.
• Teams running model safety programmes, using NeMo Auditor to probe for unsafe-prompt vulnerabilities and parallel rails to keep guardrail overhead off the critical path.
• Teams that need local inference rather than a hosted endpoint, following notebooks/08_RAG_Langchain_with_Local_NIM.ipynb for a locally deployed NIM alongside LangChain.
• Data and platform engineers working with large datasets who need GPU-accelerated knowledge graph construction through RAPIDS.
• Vision and media teams building stream monitoring or natural-language image retrieval on Vision NIMs, in addition to text-based RAG work.

Getting started

Clone the repository recursively with git clone https://github.com/nvidia/GenerativeAIExamples --recurse-submodules, which places the vision workflows at GenerativeAIExamples/vision_workflows. From there the workflows are run as Jupyter notebooks, with a "Try it Now!" section in the README pointing at the entry points.

How it compares

Among the tools named in this repository, LangChain sits at the orchestration layer and the NVIDIA serving stack — TensorRT-LLM, Triton Inference Server and NIM microservices — sits at the inference layer. GenerativeAIExamples does not compete with either; it is the glue that shows LangChain code calling NeMo Retriever NIM endpoints and consuming NIM-served models locally, with RAPIDS covering the data-processing side.

When to use it — and when not to

A self-hoster must operate the full NVIDIA stack the examples assume: GPU-accelerated infrastructure, NIM microservices, the NeMo Microservices components and the model weights behind them, none of which the repository installs for you. Teams without NVIDIA hardware, or teams deliberately keeping a vendor-neutral inference stack, should look elsewhere. The repository is also notebook-centric rather than package-centric, so there is no pip install entry point, and much of the prose documentation lives in linked blog posts and subdirectory READMEs rather than in one place; with 84 open issues, expect to read the code as the primary reference.

git clone https://github.com/nvidia/GenerativeAIExamples --recurse-submodules