Enum Normalization 

pub enum Normalization {
    PerKind,
    Global,
}

v2.6: how raw_relevance becomes the relevance term of the composite score.

Per-kind (the rule through v2.5)

Each kind is normalized against the best raw_relevance of that kind. The consequence is that the top memory and the top repo_file both score relevance = 1.0 no matter how good either actually is: on a query where memory has the answer and no file is relevant, the best of the irrelevant files is still promoted to a perfect relevance and competes for budget on the strength of the other three score terms alone.

That is the distortion the lexical and semantic floors exist to compensate for — they prune the weak candidate before normalization can flatter it. A floor is a blunt instrument for this: it is a fixed threshold standing in for a comparison the normalizer could just make.

Global

One max over all candidates, so relevance means the same thing across kinds and a candidate that is merely the best of a bad kind keeps a low relevance. Nothing else in the pipeline changes.

Which one runs

Selectable, defaulting to per_kind. Global normalization is the more principled rule and it is still not the default here, for the same reason RRF is not: the house rule is that a ranking change ships with a measurement on a real corpus, and an argument from first principles is not one. [broker] normalization and the per-request override are how a corpus gets to settle it.

Variants

PerKind

Normalize within each capsule kind. Kimetsu’s behaviour through v2.5.

Global

Normalize against a single max over every candidate.

Implementations

impl Normalization

pub fn from_config(value: &str) -> Self

Parse from config. Unknown values fall back to the default, matching how [broker] fusion treats an unrecognized rule — a typo in a config file must not silently change ranking.

Trait Implementations

impl Clone for Normalization

fn clone(&self) -> Normalization

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for Normalization

impl Debug for Normalization

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

Formats the value using the given formatter.

impl Default for Normalization

fn default() -> Normalization

Returns the “default value” for a type.

impl Eq for Normalization

impl PartialEq for Normalization

fn eq(&self, other: &Normalization) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for Normalization