Definition
Proprietary, domain-specific, or real-time data retrieved from non-parametric sources (e.g., vector stores, APIs) during inference to supplement an LLM's static training weights. It serves as the authoritative factual basis for grounding AI responses, mitigating hallucinations by decoupling logic from storage.
Distinguishes between a model's internal 'weights' and the external 'documents' it accesses at runtime.
"A lawyer (LLM) using a specialized law library (External Knowledge) to reference specific statutes instead of relying solely on their law school memory."
- Retrieval-Augmented Generation (RAG)(Implementation Framework)
- Parametric Memory(Theoretical Opposite)
- Grounding(Functional Objective)
- Vector Store(Persistence Component)
Conceptual Overview
Proprietary, domain-specific, or real-time data retrieved from non-parametric sources (e.g., vector stores, APIs) during inference to supplement an LLM's static training weights. It serves as the authoritative factual basis for grounding AI responses, mitigating hallucinations by decoupling logic from storage.
Disambiguation
Distinguishes between a model's internal 'weights' and the external 'documents' it accesses at runtime.
Visual Analog
A lawyer (LLM) using a specialized law library (External Knowledge) to reference specific statutes instead of relying solely on their law school memory.