Grounding

Grounding addresses one of the most persistent challenges in deploying LLMs: their tendency to generate plausible-sounding but factually incorrect information. A grounded model ties its responses to specific, verifiable sources — documents, databases, APIs, or search results — rather than relying solely on patterns learned during training.

The most common grounding technique is Retrieval-Augmented Generation (RAG), where relevant documents are fetched and injected into the model's context before generation. Other techniques include web search integration, database query tools, and citation-based generation where the model must quote its sources.

Grounding is especially critical in regulated industries like healthcare, finance, and law where factual accuracy is non-negotiable. Properly grounded AI systems can cite specific paragraphs, product documentation, or regulatory clauses, making their outputs auditable and trustworthy.