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
//! Nemotron reads a thread's deadline facts; the domain resolves their dates.
use super::temporal::require_explicit_fields;
use crate::domain::InputError;
use crate::domain::temporal::thread_due_date::{
ThreadDueDateRequest, ThreadFactExtractor, ThreadFacts,
};
use agent::structured_output::{DynamicSchema, validate_schema};
use agent::{Message, complete_with_history_and_params};
use ai_usage::{UsageContext, UsageRecorder};
use schemars::generate::SchemaSettings;
use serde_json::{Value, json};
use std::sync::Arc;
#[cfg(test)]
mod test;
const SCHEMA_NAME: &str = "ThreadDeadlineFacts";
/// Deadline facts through the shared model router and usage recorder.
pub struct AgentThreadFactExtractor {
recorder: Arc<dyn UsageRecorder>,
}
impl AgentThreadFactExtractor {
/// Attribute completed requests to the supplied recorder.
pub fn new(recorder: Arc<dyn UsageRecorder>) -> Self {
Self { recorder }
}
/// Actual provider schema.
pub fn schema() -> Value {
let settings = SchemaSettings::draft2020_12().with(|s| {
s.meta_schema = None;
s.inline_subschemas = true;
});
let mut schema = settings
.into_generator()
.into_root_schema_for::<ThreadFacts>()
.to_value();
require_explicit_fields(&mut schema);
schema
}
/// Exact production instructions, including the schema read by the model. Mail
/// never reaches them: it travels in the user message, as data.
pub fn prompt(schema: &Value) -> String {
format!(
"{}\n{}\n{schema}",
include_str!("../../prompts/temporal_thread_prompt.txt").trim_end(),
include_str!("../../prompts/temporal_output_prompt.txt").trim_end(),
)
}
fn model_input(request: &ThreadDueDateRequest) -> String {
let messages = request
.messages
.iter()
.enumerate()
.map(|(number, message)| {
json!({
"message": number,
"written_on": message.written_on,
"from_owner": message.from_owner,
"text": message.text,
})
})
.collect::<Vec<_>>();
json!({"subject": request.subject, "messages": messages}).to_string()
}
}
impl ThreadFactExtractor for AgentThreadFactExtractor {
#[tracing::instrument(
name = "universal_input.thread_facts",
skip_all,
err,
fields(gen_ai.request.model = super::EXTRACTION_MODEL)
)]
async fn extract_thread_facts(
&self,
request: &ThreadDueDateRequest,
usage: UsageContext,
) -> Result<ThreadFacts, InputError> {
let schema = Self::schema();
let system = Self::prompt(&schema);
let validator = validate_schema(&DynamicSchema {
schema: schema.clone(),
name: SCHEMA_NAME.into(),
description: None,
})
.map_err(|_| InputError::Extraction)?;
let result = complete_with_history_and_params(
super::EXTRACTION_MODEL,
&system,
vec![Message::user(Self::model_input(request))],
Some(json!({
"reasoning_effort": "none",
"temperature": 0,
"max_tokens": 1000,
"response_format": {"type": "json_schema", "json_schema": {
"name": SCHEMA_NAME, "strict": true, "schema": schema
}}
})),
self.recorder.as_ref(),
usage,
)
.await
.map_err(|_| InputError::Extraction)?;
let value: Value = serde_json::from_str(&result).map_err(|_| InputError::Extraction)?;
validator
.validate(&value)
.map_err(|_| InputError::Extraction)?;
serde_json::from_value(value).map_err(|_| InputError::Extraction)
}
}