Struct Engine 

pub struct Engine { /* private fields */ }

Implementations

impl Engine

pub fn new(host: Arc<dyn ModelHost>, keep_alive: Option<String>) -> Self

pub fn max_options(&self) -> usize

The largest option count this engine can answer, given what its host reports.

pub fn host_name(&self) -> &str

pub fn validate(&self, request: &SystemOneRequest) -> Result<(), EngineError>

Reject a request that cannot be answered, before any model is loaded.

pub async fn answer( &self, state: &str, question: &Question, model: &str, calibration: Calibration, ) -> Result<(Answer, u32), EngineError>

Ask one question and type its answer. The u32 is the prompt’s token count.

pub async fn answer_with_distribution( &self, state: &str, question: &Question, model: &str, calibration: Calibration, ) -> Result<(Answer, FirstTokenDistribution), EngineError>

Engine::answer, also handing back the distribution the answer was read from.

The answer renormalises over the offered labels, so it cannot say whether the model’s most likely token was a label at all. The distribution can, which is what the benchmark measures fidelity by.