AI Glossary / Grounding
⚙️ Technical
Grounding
Basing the model's response on specific sources instead of "memory" from training.
Grounding is basing the model's response on specific, verifiable material: provided documents, search results, data from tools. A "grounded" model says what follows from sources - not what sounds plausible. In practice, grounding is achieved through RAG, web search, or strict instructions like "answer only based on the attached text". It's the most effective weapon against hallucinations and the foundation of applications where errors cost: law, medicine, finance.
Related concepts