Source Attribution

The ability to link a specific claim to a precise text span within a source document, rather than citing an entire article. This requires the model to perform evidence extraction at the sentence or passage level.

• Enables verification of individual facts without reading full documents
• Relies on Faithfulness Metrics to ensure the cited span actually supports the claim
• Contrast with document-level citation, which shifts the verification burden to the user

A complete, auditable record of how information traveled from its origin to the final output. This includes the retrieval query, the ranked document list, the re-ranking decisions, and the final selected passage.

• Supports debugging when a model cites an incorrect or outdated source
• Enables Citation Precision audits by tracing each claim back through the pipeline
• Essential for regulated industries requiring full algorithmic accountability

The model should communicate its certainty about both the retrieved evidence and the generated claim. A well-calibrated system distinguishes between high-confidence citations from authoritative sources and low-confidence inferences.

• Integrates Entailment Scoring to quantify how strongly evidence supports a claim
• Uses Confidence Scores to flag statements requiring human review
• Prevents overconfident misrepresentation of weakly supported facts

Rather than relying on a single retrieved passage, robust attribution systems cross-reference claims across multiple independent sources. This mirrors journalistic fact-checking practices.

• Reduces risk of citing an erroneous or biased single source
• Enables Contradiction Detection when sources disagree
• Supports Multi-Hop Reasoning by synthesizing evidence from disparate documents
• Increases user trust through demonstrated consensus

An advanced capability where the model actively critiques its own attributions. Frameworks like Self-RAG and Chain-of-Verification (CoVe) implement this by generating verification questions and re-checking cited sources.

• The model asks: 'Does this cited passage actually state this claim?'
• Triggers Corrective RAG (CRAG) loops when initial retrieval is insufficient
• Transforms attribution from a one-time action into an iterative quality control process

Citations must be presented in a machine-readable, standardized format that enables downstream processing. This includes explicit mapping between claim IDs and source identifiers.

• Supports automated Citation Precision and Recall evaluation
• Enables integration with reference managers and knowledge bases
• Formats include inline numerical references, footnote-style links, or structured JSON output with explicit claim_id to source_span mappings

A metric evaluating the accuracy of a model's citations, measuring the proportion of generated statements with a cited source that are fully supported by that specific source.

• Key distinction: Differs from retrieval precision by evaluating the generation step, not the retrieval step
• A model may retrieve the correct document but cite it for a claim it doesn't support
• Calculated as: (Number of correctly attributed claims) / (Total number of cited claims)
• Essential for auditing faithfulness in RAG pipelines

An evaluation score that measures the degree to which a generated text is factually consistent with and can be directly inferred from the provided source document or context.

• Core principle: No external information, no fabrication, no omission of contradictions
• Often implemented using Natural Language Inference (NLI) models as automated judges
• A faithful output may still be a poor answer if the source is irrelevant, so pair with relevance metrics
• Directly combats hallucination by enforcing strict entailment from evidence

The task of automatically identifying and isolating specific text spans from a source document that directly support or refute a given claim or query.

• Input: A claim and a source document. Output: A set of character-offset spans
• Forms the backbone of automated fact-checking automation pipelines
• Modern approaches use reading comprehension models fine-tuned on datasets like FEVER
• Critical for generating human-auditable justifications in high-stakes domains like medical and legal AI

A method for reducing hallucination where a language model first drafts a response, then generates a series of independent verification questions to fact-check its own initial output.

• Four-stage process: 1. Draft response, 2. Plan verifications, 3. Execute verifications independently, 4. Produce final corrected answer
• Prevents the model from simply reinforcing its own errors by decoupling verification from generation
• Significantly reduces hallucination rate without external retrieval, using only the model's internal knowledge
• Complements RAG by adding a self-critique layer even when sources are provided

The process of using a Natural Language Inference (NLI) model to calculate a probability score indicating whether a given evidence text logically implies a target claim.

• Output classes: Entailment (supports), Contradiction (refutes), Neutral (unrelated)
• Serves as the algorithmic engine behind faithfulness metrics and automated fact-checking
• Modern NLI models like DeBERTa fine-tuned on MNLI achieve near-human performance
• Used in production to gate AI outputs: if entailment score < threshold, suppress or flag the claim

A framework that trains a language model to adaptively retrieve passages on-demand during generation and to critique its own output using special reflection tokens.

• Tokens include: Retrieve, IsRel, IsSup, IsUse, each triggering a specific behavior
• Unlike standard RAG, the model decides when to retrieve, not just what to retrieve
• The IsSup token directly evaluates whether a cited passage supports the generated claim, enabling real-time citation precision checks
• Represents a convergence of agentic RAG and source attribution into a single trained system