Enum CandidateSpec 

pub enum CandidateSpec<'a> {
    ByKey,
    Scalar {
        field: &'a str,
    },
    Tokens {
        field: &'a str,
    },
    NumericBucket {
        field: &'a str,
        tolerance: f64,
    },
    GeoGrid {
        field: &'a str,
        km: f64,
    },
    ScanAll {
        field: &'a str,
    },
    VectorClusters {
        field: &'a str,
        min: f64,
    },
    Hnsw {
        field: &'a str,
        k: usize,
    },
    Union(Vec<CandidateSpec<'a>>),
    Intersect(Vec<CandidateSpec<'a>>),
}

Variants

ByKey

Scalar

Fields

field: &'a str

Tokens

Fields

field: &'a str

NumericBucket

Fields

field: &'a str

tolerance: f64

GeoGrid

Fields

field: &'a str

km: f64

ScanAll

Fields

field: &'a str

VectorClusters

IVF-Flat approximate candidate selection (legacy; still supported as direct fallback — no longer the default for approximate: true).

k-means fitted over the indexed side’s vectors; candidates are members of the P = max(1, ceil(k/16)) nearest centroids to the query vector. NOT a superset of true positives — recall floor governs correctness.

Fields

field: &'a str

min: f64

Hnsw

HNSW approximate candidate selection (default for approximate: true).

Returns the k nearest vectors by cosine similarity from the in-tree HNSW graph. Falls back to returning all tracked nodes when the graph has no entry point (e.g. before any node is inserted, or when used without calling init_hnsw).

Fields

field: &'a str

k: usize

Number of approximate candidates to return; typically max(max_edges, 64) from the owning RuleDef.

Union(Vec<CandidateSpec<'a>>)

Union of multiple candidate specs, used for Any predicates.

Each branch of the Any predicate contributes its own candidate set (key index, token index, numeric bucket, etc.); the resulting candidate set is their union. Insert and remove recurse into every child spec so the index stays coherent for all branches simultaneously.

Intersect(Vec<CandidateSpec<'a>>)

Intersection of multiple candidate specs, used for All predicates.

Each conjunct contributes its own candidate set; the result is their intersection (empty child → empty). ScanAll children are skipped at probe time (they are the universe); if every child is ScanAll, the spec stays a full scan. Insert/remove recurse into every child.

Trait Implementations

impl<'a> Debug for CandidateSpec<'a>

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

Formats the value using the given formatter.