Struct GnnEncoder 

pub struct GnnEncoder { /* private fields */ }

Multi-layer mean-aggregation GNN encoder.

After calling GnnEncoder::fit on a set of RDF triples the encoder can produce a fixed-dimensional embedding for any entity seen during training via GnnEncoder::embed_entity. Unknown entities return a zero vector.

Implementations

impl GnnEncoder

pub fn new(config: GnnEncoderConfig) -> Self

Create a new, untrained encoder.

pub fn fit( &mut self, triples: &[(String, String, String)], ) -> Result<(), GraphRAGError>

Fit the encoder to the provided RDF triples.

Builds the adjacency graph, initialises embeddings and weight matrices, then runs stochastic gradient descent with a margin-based link-prediction loss.

pub fn embed_entity(&self, entity: &str) -> Vec<f64>

Return the embedding vector for a named entity. Returns a zero vector of length hidden_dim if the entity is unknown.

pub fn mean_aggregate(embeddings: &[&Vec<f64>]) -> Vec<f64>

Compute the mean embedding of a non-empty slice of embedding vectors.

pub fn relu_and_normalize(v: &mut [f64])

Apply ReLU activation then L2-normalise the vector in-place. If the L2 norm is near zero the vector is left unchanged.

pub fn margin_loss( pos_s: &[f64], pos_o: &[f64], neg_o: &[f64], margin: f64, ) -> f64

Triplet margin loss: max(0, d(s, o) - d(s, neg) + margin) where d is squared Euclidean distance.