Struct NliClassifier 

pub struct NliClassifier { /* private fields */ }

Zero-shot text classifier using NLI entailment scoring.

Holds a DeBERTa-v3-xsmall ONNX session and its tokenizer, each behind a Mutex so the classifier is Send + Sync and can be shared via Arc. The model is downloaded on first construction and cached locally.

Implementations

impl NliClassifier

pub fn new(config: NliConfig) -> Result<Self, NliError>

Creates a classifier from config, downloading the model if not cached.

This is synchronous and blocks on the HuggingFace download and ONNX session creation — mirroring crate::embedding::OnnxEmbedding::new. Call it from a blocking context (e.g. tokio::task::spawn_blocking) when constructing from async code. Pass NliConfig::default for the model memoir ships with.

Errors

Returns NliError::Download if the model or tokenizer cannot be fetched, NliError::ModelLoad if the ONNX session cannot be built, and NliError::TokenizerLoad if the tokenizer cannot be parsed.

pub fn execution_provider(&self) -> ExecutionProvider

Returns the execution provider this classifier resolved to at load time.

pub fn classify( &self, text: &str, labels: &[&str], hypothesis_template: &str, ) -> Result<Vec<ScoredLabel>, NliError>

Classifies text against labels using a hypothesis template.

The template must contain {}, which is replaced with each label: with template "This text is about {}." and label "a decision", the hypothesis is "This text is about a decision.". Returns the labels sorted by entailment score, highest first. An empty labels slice returns an empty vector.

Errors

Returns NliError::Inference if tokenization or model inference fails for any label.

Trait Implementations

impl Debug for NliClassifier

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.