Agentic RAG - Web Search with Accuracy and Hallucination Controls
An NVIDIA AI Workbench example that adds an agentic layer to a RAG pipeline, routing to live web search when retrieved context is thin and evaluating each response for relevance and hallucination.
This example project extends a standard retrieval-augmented-generation setup with three agentic steps: routing checks whether retrieved context is relevant to the query and falls back to live web search if it is thin, evaluation checks responses for relevance and accuracy and flags hallucinations, and iteration repeats generation and evaluation across multiple cycles.
It is built around NVIDIA AI Workbench and can run against free hosted endpoints on build.nvidia.com or be pointed at a self-hosted Ollama or NIM endpoint on your own GPUs; prompts, parameters and the UI can all be edited directly. The README documents three levels of use: using the application as-is, modifying its prompts and parameters, and self-hosting the inference GPUs.
It is distributed as a Jupyter Notebook project under the Apache-2.0 license.