Grounding
The process of anchoring an LLM's outputs to verifiable, external information sources—such as retrieved documents, database queries, or tool results—to reduce hallucination and improve factual accuracy.
Ungrounded LLMs rely entirely on knowledge encoded in their weights during training, which can be stale, incomplete, or confidently wrong. Grounding addresses this by providing a retrieval or execution mechanism: the model's claims are supported by freshly retrieved documents, live API responses, or code execution results.
The two dominant grounding patterns are RAG (retrieval-augmented generation) for knowledge grounding and tool use for action grounding. In RAG, retrieved chunks are injected into context and the model is instructed to cite them. In tool-use grounding, the model runs a calculation or queries a database instead of recalling a fact from weights.
Grounding does not eliminate hallucination—models can still misquote sources or fail to retrieve the right document—but it dramatically reduces the rate of confident factual errors. Citation-based grounding (requiring the model to quote the exact sentence supporting each claim) provides the strongest guarantee and enables automated faithfulness verification.