Struct QueryPlanner 

pub struct QueryPlanner {
    pub config: CollectionConfig,
}

Fields

config: CollectionConfig

Implementations

impl QueryPlanner

pub fn new(config: CollectionConfig) -> Self

pub fn should_use_hnsw(&self, vector_count: usize) -> bool

Returns true if the collection is large enough for HNSW to be faster than brute force (matches HnswIndex auto-build threshold).

pub fn should_pre_filter( &self, estimated_selectivity: f32, vector_count: usize, ) -> bool

Returns true when pre-filtering (metadata first, then vector scan) is beneficial.

• Selectivity < 0.1: filter is highly selective → pre-filter saves work.
• Collection > 100 000 vectors: large collections always benefit.

pub fn compute_candidate_k( &self, output_k: usize, use_hybrid: bool, has_filter: bool, ) -> usize

Compute the number of dense candidates to fetch.

Multipliers:

• hybrid + filter → ×10 (capped)
• hybrid only → ×5
• filter only → ×3
• neither → ×1

Always at least output_k, at most 10 000.

pub fn plan_search( &self, request: &SearchRequest, vector_count: usize, ) -> Result<PhysicalPlan>

Build a physical plan for a vector / text search request.

pub fn plan_sql(&self, sql: &str, vector_count: usize) -> Result<PhysicalPlan>

Build a physical plan from a SQL string.

The SQL must contain a VECTOR_SIM(col, [...]) op threshold condition. An optional scalar equality filter (AND field = value) is translated into the plan’s filter_predicate.

pub fn explain(&self, plan: &PhysicalPlan) -> String

Return a human-readable description of the physical plan.