A neural retrieval method that combines transformer models with sparse, interpretable outputs by mapping embeddings directly to vocabulary tokens.