citeformer.verify.entailment

Per-citation NLI entailment check.

For every Citation in a GenerationResult, we look up the sentence that contains the marker and check whether the cited source entails it. The output is one CitationSupport record per citation carrying the NLI entailment score and a supported boolean thresholded at the caller’s preferred value.

The premise is normally the source content (truncated on the NLI’s 512-token budget); the hypothesis is the sentence containing the citation. When the backend has populated :attr:Citation.cited_text (Anthropic’s Citations API does — see ADR-013), we use that span as the premise instead — sharper signal than the whole document, especially on long sources where the relevant assertion is buried past the 512-token horizon.

Module Contents

Functions

score_entailment

Score every citation’s entailment against its cited source.

API

citeformer.verify.entailment.score_entailment(citations: list[citeformer.core.Citation], sentence_spans: list[citeformer.verify.sentences.SentenceSpan], sources: list[citeformer.core.Source], *, nli: citeformer.verify.nli.NLIModel, threshold: float = 0.5) → list[citeformer.verify.report.CitationSupport]

Score every citation’s entailment against its cited source.

Args: citations: The citations to score. sentence_spans: Sentence spans over GenerationResult.text (from citeformer.verify.sentences.split_sentences). sources: All sources in scope; used to look up content by source_id. nli: The NLI backend. threshold: Entailment probability above which CitationSupport.         supported is True.

Returns: One CitationSupport per citation, in the same order.