Defined term
Grounding
Anchoring model output to verifiable source material to reduce hallucination.
Grounding means every claim a model produces can be traced to a retrieved source passage. Grounded outputs cite sources, refuse to answer when sources are insufficient, and flag low-confidence answers. Grounding is the strongest defense against hallucination in customer-facing AI workflows.
When it matters
When the cost of a hallucinated fact is high — legal, medical, financial, or compliance contexts. The strongest defense against fabricated answers in customer-facing AI.
Real example
A policy-explanation agent that refuses to answer unless the relevant policy passage is in the retrieved context. Every answer cites the specific section and paragraph. If retrieval fails, the agent says 'I cannot answer this from current sources' rather than guessing.
KPIs to watch
Citation rate on factual claims (100% required), groundedness eval pass rate (>95%), refusal-when-unsourced rate (should be >80% on adversarial queries).
Related terms
RAG (Retrieval-Augmented Generation)
Generation grounded in retrieved source documents rather than the model's parametric memory alone.
Hallucination
Plausible but factually incorrect output generated by an LLM with no grounding.
Guardrails
Pre and post checks that filter unsafe, off-topic, or non-compliant model outputs.
Prompt injection
An attack where user input manipulates the model into ignoring its system prompt or executing unintended actions.
See it in action
We use this every week
Book a 30-min call and we'll walk you through how Grounding shows up in a real engagement we're running.
Book a 30-min call