Glossary term
Glossary term
Foundations
Connecting model outputs to verified, authoritative sources to improve factual accuracy.
The process of basing all or part of an LLM's response on information retrieved from one or more trusted sources. For example, suppose a user prompts an LLM for today's weather forecast in Berlin. The LLM might ground the response on information it gathers from the European Centre for Medium-Range Weather Forecasts.
Retrieval-augmented generation (RAG) is a common grounding technique.
Microsoft Copilot uses Bing grounding to cite web sources for factual claims - each response includes numbered citations that users can verify, reducing unverifiable hallucinations in knowledge-work tasks.
Google Gemini's grounding with Google Search connects model responses to live search results - used by Google's AI Overviews feature to ground answers to millions of daily factual queries.
NVIDIA NeMo Guardrails' fact-checking module grounds model responses against a provided knowledge base, rejecting outputs where the stated facts cannot be attributed to a retrieved document.
Created for this library
An enterprise chatbot uses grounding by retrieving the latest policy documents before answering employee questions.
A customer support agent uses grounding by querying live order data so the response reflects current shipment status.
A field-service agent uses grounding by reading from inventory and scheduling systems before promising any service action.
Definition source: Google for Developers Machine Learning Glossary | Creative Commons Attribution 4.0 License