Groundedness Check

The check first segments generated text into atomic claims—individual, verifiable factual assertions. Each claim is isolated and evaluated independently against the source context. For example, the sentence 'Acme Corp, founded in 2012, reported $50M in revenue' yields two claims: one about the founding date and one about revenue. This granularity prevents partially supported responses from passing validation.

Uses Natural Language Inference (NLI) to determine if a source document logically entails each atomic claim. The relationship is classified as:

• Entailment: The source text directly supports the claim
• Contradiction: The source text refutes the claim
• Neutral: The source text provides insufficient evidence Only claims with entailment relationships pass the groundedness check, ensuring strict factual alignment.

Groundedness checks operate on two primary modes:

• Binary: A pass/fail determination where any unsupported claim fails the entire response, suitable for high-stakes compliance use cases
• Graded: A continuous score representing the percentage of claims with full evidential support, allowing partial credit for responses with minor unsupported elaborations Graded thresholds enable nuanced quality monitoring without blocking useful outputs.

The check strictly limits verification to the explicit context provided to the model during generation, not the model's parametric knowledge. This distinction is critical: a factually correct claim that cannot be traced to the supplied documents is flagged as ungrounded. This confinement ensures that RAG systems do not silently revert to potentially outdated or hallucinated internal knowledge.

Modern groundedness checks employ cross-encoder models that process claim-evidence pairs jointly rather than independently embedding them. This joint processing captures subtle semantic relationships—such as partial support, temporal mismatches, or numerical discrepancies—that cosine similarity on separate embeddings would miss. Cross-encoders provide higher accuracy at the cost of increased computational latency.

Groundedness checks function as streaming guardrails in production systems, intercepting generated tokens before they reach the user. When a claim fails verification, the system can:

• Truncate the response at the point of hallucination
• Regenerate with stricter decoding constraints
• Flag the output for human review while still delivering it This inline intervention prevents the propagation of unsupported information.

A quantitative evaluation score measuring the degree to which a generated statement is logically entailed by the provided source context. Unlike a binary groundedness check, faithfulness metrics produce a continuous score (e.g., 0-1) using Natural Language Inference models. This allows for nuanced quality thresholds—a response might be 0.92 faithful, triggering a revision rather than a hard rejection.

An automated evaluation step that compares a generated summary against its source material to identify contradictions, hallucinations, and unsupported inferences. Key techniques include:

• Entailment-based checking: Does the source logically imply the claim?
• Question-Answering decomposition: Generate questions from the output, answer them from the source, compare results
• Token-level alignment: Match output spans to source spans using attention weights

The underlying NLP task that powers most groundedness checks. An NLI model determines the directional logical relationship between a premise (source context) and a hypothesis (generated claim):

• Entailment: The premise guarantees the hypothesis is true
• Contradiction: The premise guarantees the hypothesis is false
• Neutral: The premise does not determine the hypothesis's truth

Modern groundedness systems fine-tune NLI models like DeBERTa on domain-specific data for higher accuracy.

A constrained text generation strategy that manipulates token probabilities during inference to favor words explicitly supported by provided evidence. Rather than checking output post-hoc, grounded decoding prevents hallucinations at generation time by:

• Masking tokens that would introduce unsupported entities
• Boosting probabilities of tokens present in the source context
• Applying prefix-tree constraints derived from retrieved documents

This shifts grounding from a verification step to a generation constraint.

The process of identifying and linking specific spans of generated text to the exact source documents that support them. While groundedness checks verify factual support, citation attribution provides the evidentiary trail. Techniques include:

• Post-hoc NLI alignment: Score each sentence against candidate sources
• Attention-based tracing: Track which retrieved chunks influenced each token
• Inline citation generation: Train models to emit [1], [2] markers during decoding

A self-correction technique where a model generates an initial response, then systematically drafts and answers independent verification questions to identify factual errors. The process:

1. Generate baseline response
2. Extract atomic claims from the response
3. Formulate verification questions for each claim
4. Answer questions using fresh retrieval (not parametric memory)
5. Revise the original response based on inconsistencies found

CoVe reduces hallucination rates by 20-40% in long-form generation tasks.