Grounding
Constraining LLM output to be supported by provided context or verified data.
Simple explanation
Grounding means the LLM only answers using information you've given it (retrieved docs, database results). If the context doesn't contain the answer, it says 'I don't know' instead of making things up.
Engineering explanation
Grounding is enforced via: system prompt ('Answer ONLY using the provided context'), delimiters around untrusted content, structured output with citation fields, and faithfulness evaluation (LLM-as-judge checks every claim is supported by context). Production RAG systems measure grounding with faithfulness scores and refuse below a threshold.
Example
Context: 'Refunds take 5-7 days.' → grounded answer: '5-7 business days [Source: Refunds Policy]'. No context → grounded answer: 'I don't have that information.'
When to use it
Any factual Q&A system, RAG, or domain-specific assistant. Anywhere hallucination is unacceptable.
Common misconception
Grounding is NOT the same as fine-tuning. Grounding is runtime context; fine-tuning changes weights. RAG grounds; fine-tuning specialises.