Struct Deduplicator 

pub struct Deduplicator {
    pub threshold: f64,
}

Fields

threshold: f64

Jaro-Winkler threshold for fuzzy merge (default 0.97). Higher = fewer false positives, more false negatives. Lower = more aggressive merging, risk of false positives.

Note: JW loses discriminative power on long names because the matching-fraction term dominates. is_variant_pair blocks prefix- extension collisions (e.g. FactoryImpl vs Factory), and the same-source-file constraint keeps comparisons within scope.

Implementations

impl Deduplicator

pub fn new(threshold: f64) -> Self

pub fn default_threshold() -> f64

pub fn deduplicate( &self, nodes: Vec<Node>, edges: Vec<Edge>, ) -> Result<(Vec<Node>, Vec<Edge>)>

Run dedup locally with no LLM calls. Pairs below 0.92 are left unmerged (conservative — no false positives). Call llm_tiebreaker separately if you want an LLM pass for higher recall.

pub fn llm_tiebreaker( &self, pairs: Vec<(Node, Node)>, ) -> Vec<(String, String, bool)>

Optional LLM tiebreaker pass for higher recall. Returns (src_id, tgt_id, merge) — caller should inspect and apply.

Trait Implementations

impl Default for Deduplicator

fn default() -> Self

Returns the “default value” for a type.