Grounding Score

Measures the semantic overlap between generated output and the retrieved context chunk. A high grounding score indicates the model's claims are directly entailed by the source text rather than fabricated.

• Uses Natural Language Inference (NLI) to classify relationships as entailment, contradiction, or neutral
• Compares vector embeddings via cosine similarity against the source passage
• Flags extrinsic hallucinations where the model introduces facts absent from the provided context

Evaluates whether every factual assertion in the output can be mapped to a specific span or passage in the source material. This prevents the model from making unsupported inferences.

• Precision: What fraction of generated claims have a corresponding source citation?
• Recall: What fraction of relevant source facts were correctly included?
• Automated systems use span alignment algorithms to match output tokens to document positions

Tests whether the grounding mechanism remains reliable when the retrieval system returns partially relevant or misleading documents. A robust score degrades gracefully rather than collapsing.

• Evaluates performance under distractor passages injected into the context window
• Measures resistance to indirect prompt injection attempting to override source adherence
• Critical for production systems where retrieval quality varies unpredictably

Grounding scores enable automated circuit breakers that block or flag outputs falling below a configurable confidence threshold. This prevents low-confidence generations from reaching end users.

• Typical production thresholds range from 0.7 to 0.9 depending on risk tolerance
• Below-threshold outputs can trigger fallback responses or human review queues
• Integrates with continuous compliance monitors to track grounding drift over time

Grounding can be assessed at multiple levels of resolution, from document-level relevance down to token-level attribution. Finer granularity provides more actionable diagnostics.

• Document-level: Is the retrieved document topically relevant?
• Sentence-level: Does each sentence align with at least one source passage?
• Claim-level: Can each atomic fact be verified against a specific source span?
• Claim-level grounding is the gold standard for faithfulness metrics in summarization tasks

A well-calibrated grounding score accurately reflects the true probability of factual correctness. Miscalibrated scores create a false sense of security or trigger unnecessary rejections.

• Expected Calibration Error (ECE) measures the gap between predicted confidence and observed accuracy
• Grounding scores should correlate strongly with human factuality judgments
• Regular recalibration against human-annotated benchmarks prevents score inflation

A quantitative score measuring the degree to which a generated summary or answer contains only claims directly inferable from the source document.

• Core Mechanism: Decomposes output into atomic claims and verifies each against the source.
• Contrast with Grounding Score: While Grounding Score measures support for any statement, Faithfulness specifically penalizes extrinsic hallucinations (added facts).
• Use Case: Critical for abstractive summarization where compression must not introduce new information.

A Natural Language Inference (NLI) task that determines whether a hypothesis (generated statement) logically follows from a premise (source text).

• Output Classes: Typically returns Entailment, Contradiction, or Neutral.
• Grounding Relationship: Serves as the binary classification backbone for many Grounding Score calculations.
• Technical Implementation: Often uses fine-tuned models like RoBERTa or DeBERTa trained on MultiNLI and SNLI datasets.

A threshold-based filter comparing vector embeddings of generated text against a reference source to block semantically divergent output.

• Mathematical Basis: Measures the cosine of the angle between two vectors in high-dimensional space, ranging from -1 to 1.
• Limitation: Captures topical similarity but not strict factual accuracy; a fluent contradiction can still have high cosine similarity.
• Pipeline Position: Often used as a fast, lightweight pre-filter before more computationally expensive entailment checks.

The frequency at which a language model generates factually incorrect, nonsensical, or unfaithful output not grounded in its training data or provided context.

• Measurement: Calculated as the percentage of generated sentences containing at least one hallucinated entity or relation.
• Inverse Relationship: A high Grounding Score directly correlates with a low Hallucination Rate.
• Taxonomy: Includes intrinsic hallucinations (contradicting the source) and extrinsic hallucinations (adding unverifiable information).

A system tracking the gradual shift in meaning or contextual relevance of generated content over time, alerting operators to topic divergence.

• Detection Method: Compares rolling window embeddings against a baseline centroid vector established from the original source material.
• Drift vs. Grounding: While Grounding Score checks factual anchoring at a point in time, Semantic Drift monitors longitudinal consistency across a generation session.
• Alert Thresholds: Triggers when the moving average cosine distance exceeds a predefined standard deviation.

The process of tracing the origin, movement, and transformation of data through a pipeline to verify integrity and ensure provenance.

• Grounding Dependency: A Grounding Score is only as reliable as the provenance of its source document; lineage audits prevent grounding against corrupted or unauthorized data.
• Metadata Tracking: Captures timestamps, schema changes, and checksums at each pipeline stage.
• Compliance Role: Essential for demonstrating to auditors exactly which source document grounded a specific generated claim.