Struct GraphScoringConfig 

pub struct GraphScoringConfig {
    pub weight_semantic: f64,
    pub weight_hotness: f64,
    pub weight_utility: f64,
    pub corroboration_boost: f64,
}

Scoring weights for utility-weighted semantic search.

The final score for a retrieved entity is computed as a linear combination of three signals:

score = weight_semantic * similarity
      + weight_hotness  * hotness
      + weight_utility  * utility_score

Defaults (0.45 / 0.30 / 0.25) match the original hard-coded values, so omitting the [graph.scoring] section from .recall-echo.toml produces identical behavior to pre-v3.9.0 recall-echo.

Weights are not constrained to sum to 1.0 — the scoring function does not normalize. Callers that change these should calibrate against their own retrieval outcomes; see utility-feedback-loop-spec.md in pulse-null.

Graph-expanded candidates score through the same three terms; what differs is where their similarity comes from (a parent’s similarity discounted by the edge’s effective confidence, rather than a direct measurement against the query vector). GraphScoringConfig::corroboration_boost governs the one case where the two channels meet.

Fields

weight_semantic: f64

Weight applied to cosine similarity. Default 0.45.

weight_hotness: f64

Weight applied to the recency/access hotness signal. Default 0.30.

weight_utility: f64

Weight applied to the utility score (outcome-feedback EMA). Default 0.25.

corroboration_boost: f64

How much an entity’s measured relevance is raised when the graph corroborates a semantic hit — i.e. when the same entity is reached both by the query vector and over a surviving edge from one of the expanded top hits. Default 0.05.

similarity = min(1.0, similarity * (1 + corroboration_boost * effective_confidence))

Scaled by the edge’s effective (decayed) confidence, so a stale edge corroborates weakly, and clamped at the similarity ceiling of 1.0, so no amount of corroboration can push an entity past what a perfect direct match would score on the same hotness and utility. 0.0 disables corroboration entirely.

The default is cut to the scale of the similarity distribution it perturbs, measured over four LongMemEval stores (196–1804 entities): the top-20 similarity band there is only 0.086 wide, so a boost of 0.134 would let corroboration promote an entity from the bottom of the band to the top, and structure would outrank similarity outright. 0.05 moves a corroborated entity about a third of the band — enough to break the near-ties that dominate a dense embedding space (the rank-1-to-rank-2 gap in those stores is 0.005–0.051), and not enough to overturn a decided ordering. Raise it only with retrieval numbers in hand: corroboration amplifies whatever the extractor put in the graph, including its mistakes.

Trait Implementations

impl Clone for GraphScoringConfig

fn clone(&self) -> GraphScoringConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GraphScoringConfig

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

Formats the value using the given formatter.

impl Default for GraphScoringConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for GraphScoringConfig

where GraphScoringConfig: 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 GraphScoringConfig

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

where __S: Serializer,

Serialize this value into the given Serde serializer.