Struct RetrievalEngine 

pub struct RetrievalEngine<B: Backend> { /* private fields */ }

Cross-modal retrieval engine.

Pre-computes embeddings for an indexed database of sensor recordings and/or text captions and enables fast nearest-neighbour search.

Implementations

impl<B: Backend> RetrievalEngine<B>

pub fn new(model: SensorLMModel<B>) -> Self

Create a new retrieval engine.

pub fn index_sensor( &mut self, sensor_batches: impl IntoIterator<Item = Tensor<B, 3>>, )

Pre-compute and store sensor embeddings from a batch of sensor data.

Arguments

• sensor_batches – Iterator of (B, T, C) sensor tensors.

pub fn index_text( &mut self, text_batches: impl IntoIterator<Item = (Tensor<B, 2, Int>, Tensor<B, 2, Int>)>, )

Pre-compute and store text embeddings from batches of token sequences.

pub fn sensor_to_text( &self, query_sensor: Tensor<B, 3>, top_k: usize, ) -> Vec<Vec<RetrievalResult>>

Retrieve the top-k most similar texts for each sensor query.

Arguments

• query_sensor – (Q, T, C) query sensor tensors.
• top_k – Number of results to return per query.

Returns

A vector of Q lists, each containing top_k RetrievalResults sorted by descending similarity.

Panics

Panics if no text embeddings have been indexed.

pub fn text_to_sensor( &self, query_ids: Tensor<B, 2, Int>, query_mask: Tensor<B, 2, Int>, top_k: usize, ) -> Vec<Vec<RetrievalResult>>

Retrieve the top-k most similar sensor recordings for each text query.

Panics

Panics if no sensor embeddings have been indexed.

pub fn evaluate_recall( &self, sensor: Tensor<B, 3>, ids: Tensor<B, 2, Int>, mask: Tensor<B, 2, Int>, k: usize, ) -> (f32, f32)

Evaluate Recall@k on a paired (sensor, text) evaluation set.

Assumes ground truth is the diagonal: sensor i corresponds to text i.

Arguments

• sensor – (N, T, C) evaluation sensor data.
• ids – (N, L) token IDs.
• mask – (N, L) attention mask.
• k – Recall@k.