What is data-designer-search-agents?

This repository is the home base for a hands-on workshop held at NVIDIA's GTC conference in San Jose in 2026, teaching people how to use a tool called NeMo Data Designer. Data Designer is used to build synthetic datasets, meaning data that is generated by an AI model rather than collected from the real world, which is often used to train or test other AI systems. The repo contains Jupyter notebooks meant to be run in order during the workshop. The first notebook introduces the basic pieces of a Data Designer workflow: starting seed data, sampling columns, and columns where an AI model writes text or structured answers, all through a simple question and answer example. The second notebook shows how to connect a web search tool to Data Designer using a protocol called MCP, so the AI model can pull in live information from the internet while generating data. The third notebook builds a more complete example: a data pipeline for search agents, covering how prompts are written, how they are varied, how the agent runs, and how quality is checked, similar to the kind of dataset NVIDIA uses to train its own Nemotron models. Setup uses a Python tool called uv to create an isolated environment and install dependencies listed in a project file. Workshop attendees need to add their own API keys for two services: one for running AI models through OpenRouter, and one for the Tavily web search tool used in two of the three notebooks. The workshop also expects certain data files, including a personas dataset from NVIDIA's catalog, to already be placed in a specific folder, and it is designed to run on machines with two H100 GPUs per person alongside a hosted GPT-OSS 120B model. This is workshop material rather than a general purpose library, so it is most useful to someone who attended or wants to follow along with the same exercises.