Hallucination Rate

Effective measurement requires classifying errors by their underlying mechanism, not just their surface appearance. A robust taxonomy enables targeted mitigation strategies.

• Factual Inaccuracy: Assertions contradicting verifiable world knowledge (e.g., incorrect dates, events, or physical laws).
• Source Unfaithfulness: Output that deviates from or fabricates claims not present in the provided grounding context or source document.
• Logical Incoherence: Statements that are internally contradictory or violate basic reasoning principles, often appearing as nonsensical chains of thought.
• Entity Fabrication: The generation of entirely fictitious persons, products, citations, or identifiers that have no real-world or source-based referent.

Hallucination rate must be evaluated separately for two distinct knowledge sources. Conflating them obscures the root cause of failure.

• Parametric Hallucination: Errors arising from the model's internal weights and pre-training data. This reflects the model's raw, unaided factual recall.
• Contextual Hallucination: Errors occurring despite correct information being present in the provided prompt or retrieved documents. This measures the model's attention and instruction-following fidelity.
• Measurement Implication: A high contextual hallucination rate indicates a failure in the retrieval or prompt architecture, while a high parametric rate points to gaps in the base model's knowledge.

Manual review does not scale. Production-grade measurement relies on automated, model-based evaluation pipelines that act as a judge.

• Natural Language Inference (NLI): Uses a separate model to determine if a generated hypothesis is entailed by, contradicts, or is neutral toward the source text. This forms the basis for faithfulness metrics.
• Fact Extraction and Verification: Pipelines that atomize generated text into discrete claims, then programmatically verify each claim against a trusted knowledge base or search index.
• Self-Consistency Checks: Sampling multiple outputs from the same prompt and measuring the semantic divergence between them. High variance often correlates with hallucination.

The granularity of measurement dictates the precision of the diagnostic signal. Choosing the wrong level can mask critical failure modes.

• Span-Level Annotation: Human or automated raters highlight specific text spans that contain unsupported information. This provides a fine-grained error density map.
• Token-Level Probability: Analyzing the model's softmax output probability for each generated token. A sequence of low-probability tokens can signal uncertainty and potential fabrication.
• Statement-Level Judgment: A binary classification of whether a full sentence or paragraph is supported. This is faster but less precise, often used for high-level monitoring dashboards.

Not all hallucinations are equally damaging. A mature measurement framework applies a risk-based weighting system to prioritize engineering resources.

• Category 1 (Critical): Fabricated legal, medical, or financial advice that could cause direct harm or liability. These errors receive the highest weight in aggregate scoring.
• Category 2 (Major): Incorrect factual assertions that damage user trust or brand credibility but do not pose immediate physical or financial danger.
• Category 3 (Minor): Stylistic exaggerations, minor date inaccuracies, or verbose filler that does not alter the core semantic meaning of the output.

In RAG architectures, the hallucination rate is directly tied to the quality of the retrieval step. Measurement must isolate the retriever's contribution to failure.

• Retriever Recall: The percentage of relevant documents successfully fetched from the vector store. Low recall forces the model to rely on parametric knowledge, increasing hallucination risk.
• Context Adherence Score: A metric specifically evaluating if the generator used the provided chunks. A low score with high retriever recall indicates a generator attention problem.
• Noise Robustness: The system's ability to ignore irrelevant or distracting documents injected into the context window, maintaining factual accuracy despite retrieval noise.

A quantitative metric that measures the degree to which a generated statement is supported by a specific source document or verified knowledge base. Unlike a general hallucination rate, the grounding score provides granular, citation-level fidelity. A low grounding score directly correlates with a high hallucination rate, indicating the model is generating text from its parametric memory rather than the provided context. This metric is critical for Retrieval-Augmented Generation (RAG) architectures where factual anchoring is non-negotiable.

A precision-oriented score measuring whether a generated summary or answer contains only claims that can be directly inferred from the source document. It specifically penalizes extrinsic hallucinations—the addition of plausible but unsupported facts. Key distinctions include:

• Faithfulness: No extra facts added (precision).
• Coverage: No facts omitted (recall). A high faithfulness metric directly suppresses the hallucination rate by enforcing strict entailment between output and source.

A Natural Language Inference (NLI) task that classifies the logical relationship between a premise (source text) and a hypothesis (generated text). The three classifications are:

• Entailment: Hypothesis is true given the premise.
• Contradiction: Hypothesis is false given the premise.
• Neutral: Hypothesis cannot be determined. This binary test serves as a real-time hallucination filter, blocking any output that fails to achieve entailment with the retrieval context before it reaches the end-user.

A threshold-based semantic filter that compares the vector embedding of the generated text against the embedding of the source document. The system computes the cosine of the angle between the two high-dimensional vectors. If the similarity score falls below a defined threshold (e.g., 0.85), the output is flagged as a potential hallucination and blocked. This method is computationally efficient for high-throughput pipelines but may miss nuanced factual errors that preserve semantic similarity.

A longitudinal monitoring system that tracks the gradual shift in meaning of generated content over time. Unlike instantaneous hallucination detection, this guardrail identifies subtle topic divergence where a model slowly begins generating off-topic or contextually irrelevant content. It works by:

• Embedding output batches over time windows.
• Computing the centroid drift from the original topic embedding.
• Alerting when drift exceeds a statistical threshold. This catches systemic degradation that single-point checks miss.

A metric measuring the alignment between a model's self-reported confidence probability and its actual empirical accuracy. A perfectly calibrated model that claims 90% confidence should be correct exactly 90% of the time. Poor calibration masks hallucination risk: an overconfident model delivers falsehoods with high certainty. Expected Calibration Error (ECE) quantifies this gap, enabling engineers to apply temperature scaling or Platt scaling to recalibrate outputs and surface uncertainty.