mod builder#
- module builder#
Prediction trie builder with incremental accumulator merge.
Ports the core algorithm from NAT’s
trie_builder.py: extract LLM call contexts from run records, compute 4-signal sensitivity scores with min-max normalization, update streaming accumulators at every trie node along the path, and build the finalPredictionTrieNodetree.Structs and Unions
- struct PredictionTrieBuilder#
Builds a
PredictionTrieNodetree fromRunRecords via incremental accumulator merge.Usage
let mut builder = PredictionTrieBuilder::new(Some(SensitivityConfig::default())); builder.add_run(&run1); builder.add_run(&run2); let trie = builder.build();
Implementations
- impl PredictionTrieBuilder#
Functions
- fn accumulators(&self) -> &AccumulatorState#
Returns a reference to the underlying accumulator state.
- fn add_run(&mut self, run: &RunRecord)#
Processes a single
RunRecordand updates accumulators.Extracts LLM call contexts, optionally computes sensitivity scores, and updates accumulators at every node along each call’s path.
- fn build(&self) -> PredictionTrieNode#
Constructs the prediction trie from accumulated data.
Iterates all accumulated nodes, navigates/creates the trie path, and populates predictions from the accumulators.
- fn new(sensitivity_config: Option<SensitivityConfig>) -> Self#
Creates a new builder with optional sensitivity scoring.
- fn with_accumulators(accumulators: AccumulatorState, sensitivity_config: Option<SensitivityConfig>) -> Self#
Creates a builder seeded with pre-existing accumulators.
Used by the learner pipeline to resume incremental learning from a stored
AccumulatorState.
- struct SensitivityConfig#
Configuration for auto-sensitivity scoring.
Weights and scale match NAT defaults from trie_builder.py lines 41-48.
- sensitivity_scale: u32#
Integer scale for quantized sensitivity (1…=scale).
- w_critical: f64#
Weight for the critical-path signal.
- w_fanout: f64#
Weight for the fan-out signal.
- w_position: f64#
Weight for the U-shaped position signal.
- w_parallel: f64#
Weight for the parallel-penalty signal.
Traits implemented
- impl Default for SensitivityConfig#