Struct RelationClassifier 

pub struct RelationClassifier { /* private fields */ }

Embedding-based relation classifier (logistic regression on MiniLM embeddings).

Implementations

impl RelationClassifier

pub fn new(model_path: &Path) -> Result<Self, RelError>

Load the ONNX model from disk.

The model expects a single input “embedding” of shape [1, 384] and produces “logits” of shape [1, 10].

pub fn with_threshold(self, threshold: f32) -> Self

Set the confidence threshold for accepting a classification.

pub fn classify_embedding( &self, embedding: &[f32], ) -> Result<Option<(String, f32)>, RelError>

Classify a relation from a pre-computed 384-dim embedding.

Returns Some((relation_type, confidence)) if a relation is detected, or None if the “none” class wins or confidence is below threshold.

pub fn classify_batch<F>( &self, text: &str, entities: &[ExtractedEntity], embed_fn: &F, ) -> Result<Vec<ExtractedRelation>, RelError>

where F: Fn(&str) -> Result<Vec<f32>, RelError>,

Classify all unique entity pairs using entity-pair embeddings.

For each pair, generates 3 embeddings (head_ctx, tail_ctx, pair_ctx) and concatenates them into a 1152-dim vector for classification. embed_fn takes text and returns a 384-dim vector.