Grounding

An ungrounded LLM generates text based entirely on patterns in its training data. It has no connection to what's actually true right now, in your documents, or about your specific situation. Grounding is how you fix that: you give the model real, verified information as part of its context, and ask it to reason from that information rather than from its training alone.

The most common grounding mechanism is RAG: before generating a response, you retrieve relevant documents from a knowledge base and include them in the prompt. The model is then instructed to base its answer on those sources. This dramatically reduces hallucination on factual questions because the model has accurate information to work with instead of guessing.

Grounding also shows up in agent contexts: an agent that can query a database, call an API, or search the web in real time is grounding its decisions in current reality rather than stale training data. This is why tool use is sometimes framed as grounding. The underlying principle is the same: the more the model's outputs are anchored in verified real-world information, the more trustworthy those outputs are.