This repository provides reference workflows and examples for building generative AI applications, specifically focusing on Retrieval Augmented Generation (RAG), agentic workflows, and model fine-tuning. It targets developers looking to integrate NVIDIA's accelerated infrastructure and microservices (NIMs) into their AI development stack, offering a streamlined path to leverage NVIDIA's ecosystem for enhanced performance and functionality.

How It Works

The project showcases end-to-end implementations using NVIDIA NeMo microservices, including components for data handling, knowledge graphs, retrieval, and guardrails. It emphasizes integration with popular frameworks like LangChain and LlamaIndex, demonstrating how to build RAG pipelines, agentic workflows, and tool-calling capabilities. The architecture leverages NVIDIA's accelerated computing and NIMs for efficient, scalable AI deployments, both in cloud and on-premises environments.

Quick Start & Requirements

• Install: Clone the repository recursively: git clone https://github.com/nvidia/GenerativeAIExamples --recurse-submodules
• Prerequisites: NVIDIA API Key (obtainable from NVIDIA API Catalog), Docker.
• Demo: docker compose up -d --build in GenerativeAIExamples/RAG/examples/basic_rag/langchain/ to run a basic RAG pipeline. Access at https://localhost:8090/.
• Documentation: Getting Started

Highlighted Details

• Demonstrates end-to-end Data Flywheel implementations using NeMo Microservices.
• Features GPU-accelerated pipelines for knowledge graph creation and querying with RAG.
• Provides examples for agentic workflows with Llama 3.1 and NeMo Retriever NIMs.
• Includes Vision NIM workflows for video stream monitoring, image search, and multimodal pipelines.

Maintenance & Community

The project is actively maintained by NVIDIA and encourages community contributions via GitHub issues and pull requests. Links to community examples and notebooks are available.

Licensing & Compatibility

The repository's code is typically licensed under permissive licenses (e.g., Apache 2.0), but specific components or dependencies might have different licensing terms. Users should verify compatibility for commercial use.

Limitations & Caveats

Some examples utilize preview NIM endpoints on the NVIDIA API Catalog, requiring an API key. On-premises deployment options are available but may involve additional setup. The recursive submodule clone is necessary for certain workflows, like Vision NIMs.