Struct SearchParams 

pub struct SearchParams {
    pub top_k: usize,
    pub n_probe: usize,
    pub n_candidate_docs: Option<usize>,
    pub centroid_score_threshold: Option<f32>,
}

Tunable knobs for a single search call.

Fields

top_k: usize

Number of top-scoring documents to return.

n_probe: usize

Number of top dot-product centroids each query token probes.

n_candidate_docs: Option<usize>

Maximum number of candidates surviving PLAID’s centroid-interaction stage that proceed to full decode + exact MaxSim. None disables centroid interaction and sends every probed candidate straight to decode — correct but slower for large corpora.

centroid_score_threshold: Option<f32>

Minimum per-centroid score (max dot product against any query token) required for a centroid to contribute to candidate generation and centroid interaction. Centroids below this threshold are treated as unreachable for this query. None disables pruning, matching the legacy probe behaviour.

Implementations

impl SearchParams

pub const fn paper_defaults(top_k: usize) -> Self

Defaults from Table 2 of the PLAID paper (Santhanam et al., 2022), bucketed by requested top_k.

top_k bucket	nprobe	t_cs	ndocs
≤ 10	1	0.5	256
≤ 100	2	0.45	1024
> 100	4	0.4	4096

The paper derives these empirically on MS MARCO and LoTTE; they trade a small recall hit for large latency wins versus an exhaustive probe. ndocs is always at least 4 * top_k so Stage 3’s ndocs/4 shortlist can still return enough candidates for the caller’s requested result count.

Trait Implementations

impl Clone for SearchParams

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for SearchParams

impl Debug for SearchParams

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

Formats the value using the given formatter.