Grounding Score

Also known as factual consistency, this is the core component of grounding. It measures the proportion of claims in the generated answer that can be directly supported by statements in the provided source context.

• Evaluation Method: Typically involves using a Natural Language Inference (NLI) model or a fine-tuned LLM judge to classify each atomic statement in the answer as entailed, contradicted, or not extractable from the context.
• Key Distinction: This metric is reference-free; it does not require a ground truth answer, only the source passages provided to the model. A high faithfulness score directly indicates a low hallucination rate.

This component evaluates the technical precision of attribution, ensuring the model not only uses the source but correctly cites it. It breaks down into two complementary metrics:

• Source Citation Recall: The proportion of source-derived statements in the answer that are correctly attributed to their originating document(s). Missed citations lower this score.
• Source Citation Precision: The proportion of citations provided in the answer that are accurate and point to a source that genuinely supports the adjacent claim. Incorrect or "fabricated" citations lower this score.

Together, they ensure the answer is auditably grounded, allowing a human or system to verify every claim.

Grounding assumes the provided context is itself relevant. This component indirectly impacts the score by assessing the quality of the retrieved passages used for generation.

• Context Relevance: Measures how pertinent the retrieved text chunks are to the query. Irrelevant context makes faithful generation impossible, capping the potential grounding score.
• Context Density: Evaluates how much of the provided context is actually utilized in the final answer. Excess, unused "noise" in the context can confuse the model and is a signal of poor retrieval precision.

A high grounding score requires that the answer faithfulness component operates on high-utility source material.

While faithfulness checks against provided sources, correctness checks against an objective ground truth. This is a stricter, composite measure.

• Relationship to Grounding: An answer can be perfectly faithful to its provided sources (high grounding score) but still incorrect if the sources themselves are wrong. Therefore, correctness is the ultimate validation of a RAG system's end-to-end accuracy.
• Measurement: Often calculated using metrics like F1 Score (token overlap) or BERTScore (semantic similarity) between the generated answer and a verified reference answer. It incorporates elements of answer relevance and factual accuracy.

Grounding scores are typically computed automatically using one of two primary techniques:

• Natural Language Inference (NLI) Models: Specialized, smaller models (e.g., DeBERTa fine-tuned on MNLI) are used to classify the relationship (entailment/contradiction/neutral) between an answer sentence and a context sentence. This is highly scalable and deterministic.
• LLM-as-a-Judge: A powerful LLM (like GPT-4) is prompted to evaluate faithfulness or generate a verifiability score based on the context and answer. This is more flexible for complex reasoning but less consistent and more expensive.

Frameworks like RAGAS and TruLens implement these methods to produce normalized grounding scores.

The Grounding Score is a critical pillar within holistic RAG assessment suites. It is often one input into a higher-level composite metric, such as a RAG Score or Answer Correctness score.

• Framework Integration: In RAGAS, it is represented by the faithfulness metric. In TruLens, it is captured by the GroundTruth or Context Relevance evals within a feedback function.
• Operational Use: It serves as a key performance indicator (KPI) for:
  • Tuning retrieval parameters to improve context quality.
  • Prompt engineering to encourage citation.
  • Monitoring production systems for drift into increased hallucination.

It transforms the qualitative concept of "factualness" into a quantitative, actionable engineering metric.

Answer Faithfulness measures the extent to which a generated answer is factually consistent with and supported by the provided source context. It is a direct precursor to Grounding Score.

• Core Focus: Factual consistency between the answer and the source.
• Evaluation Method: Typically involves cross-referencing claims in the generated answer against the retrieved context to identify contradictions or unsupported statements.
• Key Difference from Grounding: Faithfulness is binary (supported/not supported), while Grounding Score quantifies the degree of support.

Context Relevance assesses the pertinence and utility of the retrieved text passages for answering the specific query. Poor context relevance directly limits potential grounding.

• Core Focus: Quality of the retrieved information.
• Impact on Grounding: If retrieved documents are irrelevant, the model cannot produce a well-grounded answer, regardless of its generation capabilities.
• Evaluation Method: Measures if retrieved passages contain information necessary to answer the query, often using LLM-as-a-judge or entailment models.

These metrics evaluate the accuracy and completeness of a system's attribution of generated statements to specific source documents.

• Source Citation Precision: The proportion of citations in an answer that correctly point to the source of the stated information.
• Source Citation Recall: The proportion of source statements or facts used in an answer that are correctly attributed.
• Relationship to Grounding: High-precision citations are a strong signal of good grounding, as they demonstrate traceability. Grounding Score may incorporate or correlate with these metrics.

Hallucination Rate quantifies the frequency with which a generative model produces factually incorrect or unsupported statements not present in its source data. It is the inverse of good grounding.

• Core Focus: Identifying factual errors with no basis in the context.
• Direct Antithesis: A high Grounding Score implies a low hallucination rate for sourced claims. However, hallucinations can also occur on information not in the sources (extrinsic hallucinations).
• Primary Use Case: A key reliability metric for production RAG systems.

Answer Correctness is a composite metric that evaluates a generated answer's factual accuracy against a ground truth. It often incorporates aspects of faithfulness and relevance.

• Core Focus: Alignment with an objective, verifiable standard (ground truth).
• Components: Can be decomposed into Faithfulness (to source) and Answer Relevance (to query). Grounding Score primarily informs the faithfulness component.
• Evaluation Methods: Includes token-based metrics (F1, EM) and semantic metrics (BERTScore) when a ground truth is available.