RAG Faithfulness (Critical Information)
Hallucination
Overview
The RAG Faithfulness (Critical Information) evaluation measures how closely a RAG model's generated response stays grounded in the retrieved references, on questions that target critical information. It shares the faithfulness metric with the general RAG faithfulness evaluation, but its test set is deliberately concentrated on the facts that matter most: those whose incorrect answer carries real legal, financial, safety, health, security, eligibility, or discrimination / bias risk - not mere customer inconvenience.
A low faithfulness score indicates the model frequently generates claims that go beyond what the retrieved context supports, relying on parametric knowledge or hallucination instead. On critical-information questions, such ungrounded claims are the ones most likely to cause real harm.
Metrics
Faithfulness
The fraction of model response claims that are supported (entailed) by the retrieved references (range: 0.0 to 1.0).
Faithfulness is distinct from related metrics: Hallucination measures the subset of unfaithful claims that are also factually incorrect; Self-Knowledge measures unfaithful claims that are nonetheless correct. Together, hallucination and self-knowledge partition the unfaithful claims.
Motivation
In a RAG system the retrieval step exists precisely to ground the model's output in verified context. When a model ignores the retrieved references and instead draws on parametric knowledge, the retrieved context provides no safety guarantee - the model may produce plausible but outdated, incorrect, or entirely fabricated claims, and there is no way to trace those claims back to a source.
This evaluation raises the stakes by focusing the questions on critical information. When a model is unfaithful on a trivial question the cost is inconvenience; when it is unfaithful on a legal obligation, a fee, a safety instruction, or an eligibility rule, the ungrounded claim can drive a genuinely harmful decision. Measuring faithfulness specifically on these high-risk questions surfaces the failures that matter most.
Methodology
This evaluation measures whether a model's response is supported by the retrieved references, on a QA set targeting critical information.
- Test Cases: A generator model reads your text documents and writes questions focused on critical facts - those whose incorrect answer carries legal, financial, safety, health, security, eligibility, or discrimination / bias risk - each with a ground-truth answer and supporting quotes.
- Model Response: The evaluated model receives the query, retrieves relevant references from the knowledge base, and produces a response.
- Faithfulness Scoring: The model's response is broken down into atomic factual claims. Each claim is then checked against the retrieved references to determine whether it is supported.
- Score Computation: The faithfulness score is calculated as the fraction of response claims that are supported by at least one retrieved reference.
The evaluation focuses on grounding at the claim level. A response receives a higher score when more of its factual claims are supported by the retrieved references.
Claim extraction and support checking are performed by a judge model, which is configurable through the task configuration.
Scoring
Faithfulness Scorer
Examples
Faithful response - claim grounded in references
- references:- Refund requests for cancelled flights must be submitted within 90 days of the cancellation date.
The claim is directly supported by the retrieved reference.
Unfaithful response - claim not grounded in references
- references:- Refund requests for cancelled flights must be submitted within 90 days of the cancellation date.
The claim (30 days) is not supported by the retrieved reference (which states 90 days). On a critical, deadline-bearing question this ungrounded claim could cause a user to forfeit a refund.
References
[1] Ru D, Qiu L, Hu X, Zhang T, Shi P, Chang S, Jiayang C, Wang C, Sun S, Li H, Zhang Z. Ragchecker: A fine-grained framework for diagnosing retrieval-augmented generation. Advances in Neural Information Processing Systems. 2024 Dec 16;37:21999-2027.