Struct DeltaScorer 

pub struct DeltaScorer { /* private fields */ }

Answer-impact scorer: scores a chunk by whether including it would change a downstream answer, relative to the evidence already selected.

The “would the answer change?” judgement is an injectable closure (impact_fn) so the engine core never embeds a model. The closure receives the same EvidenceContext the trait does and returns a score in [0, 1]. A typical production wiring calls an LLM: “given the answer derivable from selected, does adding candidate change it? rate 0–1”.

DeltaScorer::stub ships a deterministic marginal-novelty heuristic (high when the candidate’s token set is novel vs what is already selected, low when redundant) so tests and offline runs have a model-free default that still exhibits diminishing returns.

Implementations

impl DeltaScorer

pub fn new<F>(impact_fn: F) -> Self

where F: Fn(&EvidenceContext<'_>) -> f32 + Send + Sync + 'static,

Build a scorer from a caller-supplied answer-impact closure.

pub fn stub() -> Self

Model-free stub: marginal-novelty heuristic. The candidate’s score is the fraction of its whitespace tokens that do NOT already appear in any selected candidate, scaled by its retrieval score. A fresh chunk scores near its retrieval score; a fully-redundant chunk scores ~0 — so the budget exhibits the diminishing-returns shape an answer-impact signal should.

pub fn impact(&self, ctx: &EvidenceContext<'_>) -> f32

Score via the injected closure.

Trait Implementations

impl EvidenceScorer for DeltaScorer

fn score(&self, ctx: &EvidenceContext<'_>) -> f32

Score one candidate in [0, 1].

fn name(&self) -> &str

Stable identifier surfaced in the selection diagnostics.