Struct CompressionPredictor 

pub struct CompressionPredictor { /* private fields */ }

Compression quality predictor using linear regression with sigmoid output.

Implementations

impl CompressionPredictor

pub fn new() -> Self

Create a new model with default (untrained) weights.

pub fn from_weights(weights: CompressionModelWeights) -> Self

Create a model from saved weights.

pub fn serialize(&self) -> Result<String, Error>

Serialize current weights to JSON.

Errors

Returns an error if serialization fails.

pub fn sample_count(&self) -> u64

Return the number of training samples used.

pub fn is_cold_start(&self, min_samples: u64) -> bool

Return true if the model has fewer than min_samples training samples.

During cold start, select_ratio returns None and the caller falls back to default compression behavior.

pub fn predict(&self, features: &CompressionFeatures) -> f32

Predict probe score for the given features. Range: [0.0, 1.0].

Uses sigmoid(w^T x + b) to bound predictions.

pub fn train(&mut self, samples: &[(CompressionFeatures, f32)])

Train the model on a batch of (features, probe_score) pairs using MSE loss.

Runs EPOCHS gradient steps. sample_count is incremented by the batch size once per train() call (not per epoch) to avoid over-counting.

pub fn select_ratio( &self, floor: f32, candidate_ratios: &[f32], message_count: f32, avg_message_length: f32, tool_output_fraction: f32, ) -> Option<f32>

Find the most aggressive compression ratio that keeps predicted score >= floor.

Iterates candidate_ratios from lowest (most aggressive) to highest (least aggressive). Returns the first ratio whose predicted quality clears floor, or None if no candidate passes (caller should fall back to default behavior).

Panics

Does not panic; returns None on empty candidate list.

Trait Implementations

impl Default for CompressionPredictor

fn default() -> Self

Returns the “default value” for a type.