Definition
Factual correctness in RAG pipelines is the degree to which generated claims are logically entailed by the retrieved context. It requires balancing strict grounding (reducing hallucinations) against the model's ability to synthesize information across multiple document fragments.
Focuses on 'faithfulness' to the retrieved context rather than the LLM's internal pre-trained world knowledge.
"The Court Reporter's Transcript: ensuring the summary of the trial matches exactly what was entered into evidence, without adding outside rumors."
- Hallucination(Inverse/Failure State)
- Grounding(Prerequisite)
- Natural Language Inference (NLI)(Underlying Technique)
- Faithfulness(Component Metric)
Conceptual Overview
Factual correctness in RAG pipelines is the degree to which generated claims are logically entailed by the retrieved context. It requires balancing strict grounding (reducing hallucinations) against the model's ability to synthesize information across multiple document fragments.
Disambiguation
Focuses on 'faithfulness' to the retrieved context rather than the LLM's internal pre-trained world knowledge.
Visual Analog
The Court Reporter's Transcript: ensuring the summary of the trial matches exactly what was entered into evidence, without adding outside rumors.