Struct GraphDedupConfig 

pub struct GraphDedupConfig {
    pub certain_similarity: f64,
    pub review_similarity: f64,
    pub max_candidates: usize,
}

Similarity bands that decide when entity dedup pays for a model call.

Dedup asks one question — is this the same thing? — and that is a question about meaning, so the bands are cut on raw cosine similarity between the candidate’s abstract and an existing entity’s, never on the retrieval score (which folds in hotness and utility: a popular unrelated entity would otherwise buy a model call, and every entity gets more popular as the graph grows).

similarity >= certain_similarity   → the same entity; resolved locally
review_similarity ..< certain      → ambiguous; one model call decides
similarity <  review_similarity    → new entity; created locally

Defaults (0.92 / 0.82 / 3) are cut from the similarity distribution of a LongMemEval baseline store (192 entities, 150 sampled candidates, 750 neighbour pairs). BGE-Small puts every same-language pair in a narrow high band — median neighbour 0.75, median nearest neighbour 0.81 — so the cuts sit at its tail, not at intuitive-looking round numbers: 0.92 is the 96th percentile of pairs, where abstracts are paraphrases of each other, and 0.82 the ~78th, below which pairs are merely same-topic. Candidates averaged 1.1 neighbours above 0.82, so a cap of three bounds the worst case without binding the normal one.

Fields

certain_similarity: f64

At or above this cosine similarity the candidate is treated as the same entity and resolved without a model call. Default 0.92.

review_similarity: f64

Below this cosine similarity the candidate is treated as new and created without a model call. Default 0.82.

max_candidates: usize

How many existing entities, by similarity rank, dedup may fetch and hand to the model. Caps prompt size and comparison count so neither can grow with the graph. Default 3.

Implementations

impl GraphDedupConfig

pub fn band(&self, similarity: f64) -> DedupBand

The band a candidate’s nearest neighbour falls in.

pub fn candidate_limit(&self) -> usize

How many candidates to fetch and consider — at least one, whatever the config says, or dedup would be blind.

Trait Implementations

impl Clone for GraphDedupConfig

fn clone(&self) -> GraphDedupConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GraphDedupConfig

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

Formats the value using the given formatter.

impl Default for GraphDedupConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for GraphDedupConfig

where GraphDedupConfig: Default,

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for GraphDedupConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.