Semantic Intent Classification

Unlike rigid keyword regexes, Whisper Bot translates user queries into control commands using semantic proximity vectors generated by a local sentence-transformer.

Intent Engine Workflow

Model: all-MiniLM-L6-v2 sentence transformer compiled to ONNX format. Maps any natural language sentence into a 384-dimensional floating-point vector.
Embedding Generation: The model extracts token features and performs mean pooling over the sequence dimension to yield a single embedding vector.
Proximity Comparison: The query vector is compared against a pre-compiled database of 28 movement intents using Cosine Similarity.

Mathematical Formulation

Cosine similarity measures the angular distance between the query vector A and intent vector B:

Similarity = (A · B) / (||A|| * ||B||)

Since all embeddings in our database are pre-normalized to unit length (||A|| = ||B|| = 1), the calculation simplifies to a simple dot product:

Similarity = A · B