Struct FeedbackLoop 

pub struct FeedbackLoop { /* private fields */ }

Closed-loop feedback collector that ties together the store and reward model.

Implementations

impl FeedbackLoop

pub fn new( ring_capacity: usize, decay: f64, max_tokens: u64, max_latency_ms: u64, ) -> Self

Create a new FeedbackLoop.

• ring_capacity: number of feedbacks retained per variant in the ring buffer.
• decay: exponential smoothing factor for the reward model (0 < decay < 1).
• max_tokens: token ceiling used when normalising cost penalties.
• max_latency_ms: latency ceiling used when normalising latency penalties.

pub fn record(&self, feedback: Feedback)

Record a feedback observation.

The variant key is derived from request_id as the model name prefix (everything before the first ':'). Callers should embed the model name in request_id as "model:rest-of-id". If no ':' is found the whole request_id is used as the variant key.

pub fn best_variant(&self, model_options: &[String]) -> Option<String>

Return the variant (from model_options) with the highest average reward.

Returns None if none of the provided options have any recorded rewards.

pub fn reward_history(&self, variant: &str) -> Vec<f64>

Return the chronological reward history for a given variant.

pub fn convergence_score(&self) -> f64

Compute a convergence score as the variance of the most recent rewards across all known variants.

A low value (close to 0) indicates that rewards have stabilised.

Trait Implementations

impl Clone for FeedbackLoop

fn clone(&self) -> FeedbackLoop

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for FeedbackLoop

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.