It's a compressed database with diffuse indices. It's using probability matching rather than pattern matching. Write operations are called 'training' and 'fine-tuning'.