Grounding

Grounding is the practice of constraining a language model's responses to be based on specific retrieved evidence rather than the model's internal (parametric) knowledge. In a RAG system, grounding means instructing the LLM to answer only using the provided context documents and to explicitly state when the answer cannot be found in the available evidence.

Grounding is the primary mechanism for hallucination control in production RAG systems. Without grounding, LLMs may generate plausible-sounding but factually incorrect information. Grounding techniques include prompt instructions ("only use information from the context"), citation requirements ("cite the source for each claim"), the "I don't know" pattern ("if the answer is not in the context, say so"), and post-generation verification that checks claims against source documents.

In hybrid RAG+KG systems, grounding is strengthened by combining evidence from multiple sources: structured facts from the knowledge graph and unstructured supporting text from document retrieval. When both sources agree, confidence is high. Disagreements can be flagged for human review. This multi-source grounding produces more reliable answers than either source alone.