Source Attribution

The core mechanism that maps a generated statement to a specific span of text within a source document, not just the document as a whole. This is achieved through attention weight analysis or by prompting the model to generate inline citations. In legal AI, this means the difference between citing a 200-page contract and pointing to the exact clause, such as Section 12.3(a), that supports the analysis. This granularity enables rapid human verification and is the foundation of a trustworthy audit trail.

An architectural pattern where attribution is performed by a secondary system after the primary model generates a response. This engine decomposes the generated text into atomic factual claims and uses a high-recall retrieval model to search the source corpus for supporting evidence. A Natural Language Inference (NLI) model then classifies whether each claim is entailed by the retrieved evidence. This decoupled approach allows for rigorous verification without constraining the generative model's fluency.

A technique where the language model is fine-tuned or prompted to generate references directly within its output stream, similar to a human writing a legal memo. The model learns to append bracketed markers like [12] or (Smith v. Jones, 2023) immediately after a factual assertion. This requires a specialized training dataset where every sentence in the corpus is annotated with its canonical source identifier. The result is a seamless reading experience where the evidence is presented concurrently with the conclusion.

A suite of quantitative measures to evaluate the quality of a source attribution system. Key metrics include:

• Citation Recall: The percentage of generated claims that have a supporting citation.
• Citation Precision: The percentage of provided citations that genuinely support their associated claim, directly measuring the rate of fabricated references.
• Attribution F1 Score: The harmonic mean of recall and precision, providing a single balanced measure of attribution quality. These metrics are critical for establishing a legal AI's reliability before deployment.

A runtime guardrail that prevents the model from introducing external knowledge not present in the provided source documents. Before a token is generated, the system's logits are modified to penalize tokens that would lead to statements unsupported by the retrieved context. This is implemented by comparing the model's predicted output distribution against a constrained distribution derived solely from the source material. This technique is essential for maintaining strict evidentiary boundaries in legal reasoning, ensuring the AI does not 'fill in the gaps' with potentially hallucinated information.

A methodology for handling conflicting information across a multi-document corpus by establishing a precedence hierarchy among sources. In a legal context, this means the system is configured to prioritize a binding Supreme Court ruling over a dissenting opinion from a lower court, or a specific contract clause over a general boilerplate provision. When generating an answer, the attribution system resolves conflicts by deferring to the highest-authority source and explicitly noting the overruled reference, providing a transparent reasoning chain.

The automated process of verifying that every factual claim in a generated text is explicitly supported by the provided source document. This serves as a critical guardrail against hallucination in legal AI systems.

• Compares each generated sentence against source text
• Uses Natural Language Inference (NLI) to classify support, contradiction, or neutrality
• Produces an attribution score quantifying the percentage of grounded claims

Two complementary metrics that measure the integrity of a model's source attribution. Citation Recall quantifies the proportion of factual claims correctly supported by a citation. Citation Precision measures the proportion of provided citations that genuinely support their associated claim.

• High recall + low precision = over-citation with irrelevant sources
• Low recall + high precision = under-citation with missing authority
• Both metrics must be optimized to prevent fabricated references

A classification task that determines whether a hypothesis can be logically inferred from a premise. In source attribution pipelines, NLI is used to check if a generated statement is entailed by, contradicts, or is neutral to the source text.

• Entailment: The source text logically supports the claim
• Contradiction: The source text refutes the claim
• Neutral: The source text provides no information about the claim

A metric that quantifies the degree to which a generated statement can be directly linked to a specific segment of a source document. This ensures every legal conclusion has a verifiable provenance.

• Operates at the span level, not just document level
• Uses token-level alignment between output and source
• Enables fine-grained audit trails for regulatory compliance

A multi-stage automated system that decomposes a claim, retrieves relevant evidence from a trusted corpus, and uses an NLI model to render a verdict on the claim's veracity. This is the production architecture for source attribution.

• Claim decomposition breaks complex statements into atomic facts
• Evidence retrieval fetches candidate supporting passages
• Verdict rendering classifies each fact as supported, refuted, or unverifiable