SmartFAQs.ai
Back to Learn
Intermediate

Knowledge Conflict

Knowledge conflict occurs in RAG pipelines when the non-parametric data retrieved from an external data source contradicts the parametric knowledge embedded in the LLM's weights during pre-training. This discrepancy forces the model to choose between its internal 'world model' and the provided external context, often leading to performance degradation if not handled via prompt engineering or architectural weighting.

Definition

Knowledge conflict occurs in RAG pipelines when the non-parametric data retrieved from an external data source contradicts the parametric knowledge embedded in the LLM's weights during pre-training. This discrepancy forces the model to choose between its internal 'world model' and the provided external context, often leading to performance degradation if not handled via prompt engineering or architectural weighting.

Disambiguation

Differentiates between a model being 'wrong' (hallucination) and a model being 'conflicted' (internal training vs. external context).

Visual Metaphor

"A judge weighing a new witness statement (Retrieved Context) against their own personal memory of the event (Parametric Weights)."

Key Tools
LangChainNeMo GuardrailsGuardrails AILlamaIndex
Related Connections

Conceptual Overview

Knowledge conflict occurs in RAG pipelines when the non-parametric data retrieved from an external data source contradicts the parametric knowledge embedded in the LLM's weights during pre-training. This discrepancy forces the model to choose between its internal 'world model' and the provided external context, often leading to performance degradation if not handled via prompt engineering or architectural weighting.

Disambiguation

Differentiates between a model being 'wrong' (hallucination) and a model being 'conflicted' (internal training vs. external context).

Visual Analog

A judge weighing a new witness statement (Retrieved Context) against their own personal memory of the event (Parametric Weights).

Related Articles