Struct Faithfulness 

pub struct Faithfulness<M: BaseChatModel> { /* private fields */ }

忠实度评测器(幻觉检测):回答有多忠于参考上下文。

把 prediction 拆成原子陈述,逐条问裁判能否从 reference 推导, 通过率 = 能推导的陈述数 / 总陈述数。

Implementations

impl<M: BaseChatModel> Faithfulness<M>

pub fn new(judge: M) -> Self

创建忠实度评测器。

pub fn with_llm_split(self, v: bool) -> Self

用 LLM 拆原子陈述(默认按标点规则拆;LLM 拆能处理逗号复合句)

pub fn with_empty_score(self, score: f64) -> Self

空预测的得分:默认 0.0(没回答=不忠实),可设 1.0 表示“没编造即忠实“

pub fn with_max_context_chars(self, max: usize) -> Self

参考上下文单条传输上限(字符)。P2-5: 默认 2000,防止长参考被每条陈述重复完整塞进 prompt。

Trait Implementations

impl<M: BaseChatModel> Evaluator for Faithfulness<M>

fn eval<'life0, 'life1, 'life2, 'life3, 'async_trait>( &'life0 self, _input: &'life1 str, prediction: &'life2 str, reference: &'life3 str, ) -> Pin<Box<dyn Future<Output = Result<Score, EvalError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, 'life3: 'async_trait,

对单条预测打分

fn name(&self) -> &str

评测器名称(用于报告汇总)