# Grounding

> Grounding ties a model's output to verifiable sources — retrieved documents, tool results, citations — instead of training-data memory.

**Grounding is anchoring a model's output to verifiable evidence — retrieved documents, tool results, supplied sources — so answers come from checkable material rather than the model's training-data memory.**

It's the direct countermeasure to [hallucination](/glossary/hallucination): a model generating freely produces the most *plausible* continuation, while a grounded model is constrained to the most *supported* one. The mechanics have three parts — deliver evidence at query time ([RAG](/glossary/rag) being the workhorse delivery system), instruct the model to answer only from it (with "the sources don't say" as an explicitly allowed move), and make fidelity visible through citations, so an ungrounded claim is detectable rather than smooth.

Grounding quality is measurable — faithfulness metrics score whether answers follow from sources — which makes it an engineering property, not a vibe. The full pipeline that produces well-grounded answers is [How RAG Actually Works](/guides/concepts/how-rag-works); what breaks when answers float free of their sources is step four of the [RAG debugging checklist](/guides/troubleshooting/rag-debugging-checklist).

---

_Source: https://agentscamp.com/glossary/grounding — Term on AgentsCamp._