Files
openworker/harness/observability/src/evaluations.rs

296 lines
9.3 KiB
Rust

//! Optional, acknowledged evaluation records. Never sent through the lossy trace queue.
use anyhow::{ensure, Result};
pub const DEFAULT_MAX_PENDING: u32 = 10_000;
use serde::{Deserialize, Serialize};
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct Target {
pub session_hash: String,
pub turn_hash: String,
pub trace_id: Option<String>,
pub span_id: Option<String>,
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct Evaluator {
#[serde(default, skip_serializing_if = "Option::is_none")]
pub model: Option<String>,
#[serde(default, skip_serializing_if = "Option::is_none")]
pub usage: Option<crate::Usage>,
pub name: String,
pub version: String,
pub source: Source,
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum Source {
Human,
Code,
Model,
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
#[serde(
tag = "type",
content = "value",
rename_all = "snake_case",
deny_unknown_fields
)]
pub enum Score {
Numeric(f64),
Boolean(bool),
Categorical(String),
Text(String),
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum Status {
Completed,
Skipped,
Error,
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct Input {
pub id: String,
pub name: String,
pub evaluator: Evaluator,
pub status: Status,
pub score: Option<Score>,
pub created_ms: u64,
pub explanation: Option<String>,
pub dataset_case: Option<String>,
pub experiment: Option<String>,
/// Corrections append a new ID, preserving the original evaluation.
pub supersedes: Option<String>,
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct Evaluation {
pub version: u8,
pub deployment: String,
pub target: Target,
pub result: Input,
}
impl Evaluation {
pub fn validate(&self) -> Result<()> {
ensure!(
self.version == 1 && crate::identifier(&self.deployment),
"Unsupported evaluation schema or deployment"
);
ensure!(
crate::hex_id(&self.target.session_hash, 64)
&& crate::hex_id(&self.target.turn_hash, 64),
"Invalid evaluation scope"
);
ensure!(
self.target
.trace_id
.as_ref()
.is_none_or(|v| crate::hex_id(v, 32))
&& self
.target
.span_id
.as_ref()
.is_none_or(|v| crate::hex_id(v, 16))
&& (self.target.span_id.is_none() || self.target.trace_id.is_some()),
"Invalid evaluation trace target"
);
for id in [
&self.result.id,
&self.result.name,
&self.result.evaluator.name,
&self.result.evaluator.version,
]
.into_iter()
.chain(self.result.dataset_case.iter())
.chain(self.result.experiment.iter())
.chain(self.result.supersedes.iter())
{
ensure!(crate::label(id), "Invalid evaluation identifier");
}
ensure!(
self.result
.evaluator
.model
.as_ref()
.is_none_or(|m| crate::label(m)),
"Invalid evaluator model"
);
ensure!(
(self.result.evaluator.model.is_none() && self.result.evaluator.usage.is_none())
|| self.result.evaluator.source == Source::Model,
"Only model evaluators carry model accounting"
);
if let Some(u) = &self.result.evaluator.usage {
ensure!(
openworker_telemetry::TokenUsage {
input: u.input,
output: u.output,
cache_read: u.cache_read,
cache_write: u.cache_write,
reasoning: u.reasoning
}
.validate(),
"Invalid evaluator usage"
);
}
ensure!(
self.result.created_ms > 0 && self.result.created_ms <= 253402300799000,
"Invalid evaluation timestamp"
);
ensure!(
(self.result.status == Status::Completed) == self.result.score.is_some(),
"Only completed evaluations carry a score"
);
ensure!(
self.result.supersedes.as_ref() != Some(&self.result.id),
"An evaluation cannot supersede itself"
);
if let Some(score) = &self.result.score {
match score {
Score::Numeric(n) => ensure!(n.is_finite(), "Numeric score must be finite"),
Score::Categorical(s) => {
ensure!(!s.is_empty() && s.len() <= 256, "Invalid category")
}
Score::Text(s) => ensure!(s.len() <= 4000, "Score text exceeds limit"),
Score::Boolean(_) => {}
}
}
ensure!(
self.result
.explanation
.as_ref()
.is_none_or(|s| s.len() <= 4000),
"Evaluation explanation exceeds limit"
);
ensure!(
serde_json::to_vec(self)?.len() <= 16384,
"Evaluation exceeds record limit"
);
Ok(())
}
/// Best-effort secret redaction precedes durable storage and export.
pub fn redacted(mut self) -> Self {
if let Some(text) = &mut self.result.explanation {
*text = openworker_telemetry::redact(text);
}
if let Some(Score::Text(text) | Score::Categorical(text)) = &mut self.result.score {
*text = openworker_telemetry::redact(text);
}
self
}
}
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Row {
pub cursor: u64,
pub evaluation: Evaluation,
}
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Page {
pub results: Vec<Row>,
pub next: Option<u64>,
}
#[derive(Clone, Debug, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct Query {
pub deployment: String,
pub session_hash: String,
pub turn_hash: String,
pub after: u64,
pub limit: u16,
}
impl Query {
pub fn validate(&self) -> Result<()> {
ensure!(
crate::identifier(&self.deployment)
&& crate::hex_id(&self.session_hash, 64)
&& crate::hex_id(&self.turn_hash, 64)
&& self.after <= i64::MAX as u64
&& (1..=50).contains(&self.limit),
"Invalid evaluation query"
);
Ok(())
}
}
pub(crate) fn page(mut results: Vec<Row>, q: &Query) -> Result<Page> {
for row in &results {
row.evaluation.validate()?;
ensure!(
row.cursor > q.after
&& row.evaluation.deployment == q.deployment
&& row.evaluation.target.session_hash == q.session_hash
&& row.evaluation.target.turn_hash == q.turn_hash,
"Evaluation scope mismatch"
);
}
let more = results.len() > usize::from(q.limit);
results.truncate(q.limit.into());
let next = more.then(|| results.last().unwrap().cursor);
Ok(Page { results, next })
}
#[cfg(test)]
mod tests {
use super::*;
pub fn sample() -> Evaluation {
Evaluation {
version: 1,
deployment: "sdk".into(),
target: Target {
session_hash: format!("{:064x}", 1),
turn_hash: format!("{:064x}", 2),
trace_id: None,
span_id: None,
},
result: Input {
id: "e1".into(),
name: "correctness".into(),
evaluator: Evaluator {
name: "fixture".into(),
version: "1".into(),
source: Source::Code,
model: None,
usage: None,
},
status: Status::Completed,
score: Some(Score::Boolean(true)),
created_ms: 1,
explanation: None,
dataset_case: None,
experiment: None,
supersedes: None,
},
}
}
#[test]
fn status_and_target_cannot_misrepresent_results() {
let mut e = sample();
assert!(e.validate().is_ok());
e.result.status = Status::Error;
assert!(e.validate().is_err());
e.result.score = None;
assert!(e.validate().is_ok());
e.target.span_id = Some(format!("{:016x}", 1));
assert!(e.validate().is_err());
}
}
/// Reuses the compliance outbox lease/backoff contract; only evaluation delivery
/// is retried. No runtime action is addressed by this checkpoint.
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Delivery {
pub token: String,
pub row: Row,
}
pub fn delivery_key(deployment: &str, sink: &str, binding: &str, token: &str) -> Result<()> {
ensure!(
crate::identifier(deployment)
&& crate::identifier(sink)
&& crate::hex_id(binding, 64)
&& crate::identifier(token),
"Invalid evaluation delivery scope"
);
Ok(())
}