Definition
Grounding is the architectural process of anchoring an LLM's output to a specific, verifiable corpus of data provided in the context window, ensuring the model's claims are derived from retrieved facts rather than internal weights. It serves as the primary mechanism for mitigating hallucinations in RAG pipelines by enforcing a 'closed-book' constraint on the generation phase.
Refers to factual anchoring in external data, not electrical grounding or linguistic semantics.
"An Open-Book Exam where the student is strictly forbidden from answering using personal memory and must cite specific page numbers from the provided textbook."
- Hallucination(Target failure state)
- Context Window(Physical constraint for grounding data)
- Faithfulness(Evaluation metric for grounding)
- Source Citation(Implementation technique)
Conceptual Overview
Grounding is the architectural process of anchoring an LLM's output to a specific, verifiable corpus of data provided in the context window, ensuring the model's claims are derived from retrieved facts rather than internal weights. It serves as the primary mechanism for mitigating hallucinations in RAG pipelines by enforcing a 'closed-book' constraint on the generation phase.
Disambiguation
Refers to factual anchoring in external data, not electrical grounding or linguistic semantics.
Visual Analog
An Open-Book Exam where the student is strictly forbidden from answering using personal memory and must cite specific page numbers from the provided textbook.