Grounding Score

A Grounding Score is a quantitative metric (0-1 or percentage) representing the semantic overlap between a generated response and its source document. It typically leverages a Natural Language Inference (NLI) model or a specialized embedding similarity model to verify if a hypothesis (the generation) is entailed by the premise (the retrieved chunk). A high score indicates the model strictly adhered to the provided context, while a low score signals potential hallucination or fabrication.

Grounding is not a single algorithm but a class of techniques:

• NLI-Based Scoring: Uses models like BART or DeBERTa fine-tuned on MNLI to classify sentence pairs as 'entailment', 'contradiction', or 'neutral'. The score is the probability of entailment.
• Embedding Similarity: Calculates the cosine similarity between dense vector representations of the claim and the source text.
• Claim Decomposition: Breaks the output into atomic facts and verifies each independently against the source, often using a FActScore-style approach.

While often used interchangeably, Grounding and Faithfulness have subtle differences:

• Grounding strictly measures alignment with the retrieved external context provided in the prompt.
• Faithfulness is broader, measuring if the output is consistent with the entire input history, including system prompts and conversation turns. A response can be faithful to the conversation but ungrounded if it ignores the retrieved document.

In production RAG systems, the Grounding Score acts as a binary gatekeeper. A typical architecture involves:

1. Retrieval: Fetch top-k documents.
2. Generation: Produce a response.
3. Grounding Check: Calculate the score.
4. Conditional Logic: If the score falls below a threshold (e.g., < 0.7), the system triggers a fallback strategy, such as responding with 'I don't know' or re-running the retrieval with a different query. This is a core component of NeMo Guardrails.

Grounding Scores are not infallible:

• Semantic Blindness: High cosine similarity does not guarantee factual accuracy if the source itself is wrong.
• NLI Brittleness: NLI models struggle with negation and complex numerical reasoning.
• Granularity Mismatch: A single score for a long passage can mask a single critical hallucination. Mitigation requires sentence-level or atomic fact-level scoring for high-stakes applications.

Grounding Score and Hallucination Rate are inversely correlated but measure different things. Hallucination Rate is a human-centric evaluation of factual errors per token. Grounding Score is an automated, proxy metric. A low Grounding Score is a strong predictor of a high Hallucination Rate, making it a scalable, real-time monitoring tool for LLMOps pipelines where human review is too slow.

A core metric evaluating whether all factual claims in a generated text are supported by a source document. It measures the alignment between the output and the provided grounding context. Unlike simple n-gram overlap, modern implementations use Natural Language Inference (NLI) to classify if a hypothesis (the generation) is entailed by the premise (the source).

• Key Mechanism: NLI-based entailment scoring
• Use Case: Summarization and RAG output validation
• Limitation: Does not detect omissions or logical contradictions across multiple sentences

An automated evaluation score that determines if a generated summary or response can be logically deduced from the input source without introducing extraneous information. It typically decomposes text into atomic factual units and checks each against the source.

• Core Technique: Decompose-then-verify using NLI
• Critical Distinction: A faithful output can be incomplete but never contradictory
• Benchmark: FaithDial dataset provides human-corrected faithful responses for training

A composite metric evaluating whether a model can correctly link a generated claim to the specific segment of a source document that supports it. It is measured by two sub-metrics:

• Citation Recall: The proportion of generated claims that are supported by a cited source
• Citation Precision: The proportion of citations that actually support the corresponding claim

High attribution scores are essential for enterprise auditability and regulatory compliance in regulated industries.

A human-aligned evaluation metric that breaks a long-form generation into atomic facts and verifies each against a trusted knowledge base like Wikipedia. It calculates the percentage of supported facts, providing a granular view of factual density.

• Process: Atomic fact decomposition → per-fact verification → binary entailment decision
• Strength: Correlates strongly with human judgments of factuality
• Application: Evaluating biographies, product descriptions, and technical documentation generated by LLMs

A measure of uncertainty in language model outputs that clusters semantically equivalent generations before calculating entropy. This distinguishes between high uncertainty (the model is genuinely confused) and simple lexical variation (the model is confident but paraphrasing).

• Method: Sample multiple outputs → cluster by semantic equivalence → compute entropy over clusters
• Advantage: Detects confabulation more reliably than token-level entropy
• Implementation: Used in SelfCheckGPT for zero-resource hallucination detection

A prompting technique where an LLM first drafts a response, then generates a series of independent verification questions to fact-check its own work, and finally produces a corrected, verified answer.

• Step 1: Generate baseline response
• Step 2: Plan verification questions based on the draft
• Step 3: Execute verifications independently
• Step 4: Produce final answer incorporating verification results

This approach reduces hallucination rates by decoupling generation from verification, preventing the model from simply rationalizing its own errors.