Macro is a unified workspace for teams: email, chat, docs, tasks, agents, calls, and CRM — @-linked together with shared AI memory.
TypeScript agentaiai-agentsall-in-onecrmcrm-systememaillinearmcpmessagingnotesnotionofficerustslackslack-alternativestartupstartup-toolstypescriptworkspace
//! The metered classification use case.
use std::sync::Arc;
use ai_usage::{UsageAmount, UsageContext, UsageRecorder};
use serde_json::Value;
use super::{
models::{JevError, MAX_QUESTIONS, Probability, YesNoQuestion},
ports::{JevProvider, YesNoClassifier},
};
#[cfg(test)]
mod test;
/// Classifies through a [`JevProvider`] and records each round trip's tokens.
pub struct JevClassifier<P> {
provider: P,
recorder: Arc<dyn UsageRecorder>,
}
impl<P: JevProvider> JevClassifier<P> {
/// Build a classifier that meters usage through `recorder`.
pub fn new(provider: P, recorder: Arc<dyn UsageRecorder>) -> Self {
Self { provider, recorder }
}
}
impl<P: JevProvider> YesNoClassifier for JevClassifier<P> {
#[tracing::instrument(name = "jev.classify", skip_all, err, fields(
jev.questions = questions.len(),
ai.feature = ?usage.feature,
))]
async fn classify(
&self,
usage: UsageContext,
input: &Value,
questions: &[YesNoQuestion],
) -> Result<Vec<Probability>, JevError> {
if questions.is_empty() {
return Err(JevError::NoQuestions);
}
if questions.len() > MAX_QUESTIONS {
return Err(JevError::TooManyQuestions);
}
let evaluation = self.provider.evaluate(input, questions).await?;
// The provider billed the call even if its answers are unusable.
// Jev has no prompt cache, so every input token is uncached.
self.recorder.record(usage.into_event(
self.provider.model_id().to_owned(),
UsageAmount::Tokens {
input: evaluation.input_tokens,
output: evaluation.output_tokens,
cache_read: 0,
cache_write: 0,
},
));
if evaluation.probabilities.len() != questions.len() {
return Err(JevError::InvalidResponse);
}
Ok(evaluation.probabilities)
}
}