Struct SearchSession 

pub struct SearchSession { /* private fields */ }

Pre-loaded search session: all HNSW indexes loaded into memory once.

Useful for benchmarks and servers that issue many queries against the same snapshot. Avoids re-loading and re-deserializing indexes on every call.

Deleted rows are NOT filtered here: unlike search/search_text, this session does not load deletion vectors or equality delete files — rows removed via delete_rows/delete_where still appear in results. Use search when delete visibility matters; this type trades that for raw throughput on static snapshots (its benchmark use case).

Implementations

impl SearchSession

pub async fn load( table: &TableIdent, vector_column: &str, dim: u32, catalog: Arc<dyn CatalogProvider>, store: Arc<dyn Store>, load_raw: bool, ) -> AilakeResult<Self>

Load all indexes for the latest snapshot into memory.

Pass load_raw = true when reranking will be used (rerank_factor is Some); it reads the full parquet columns so exact distances are available without extra I/O during search_query.

pub fn shard_count(&self) -> usize

Number of loaded shards.

pub fn search_batch( &self, queries: &[Vec<f32>], config: &SearchConfig, ) -> Vec<Vec<SearchResult>>

Search multiple queries in one call.

For shards with raw vectors (Indexing or reranking): dispatches to GPU batch matmul when a CUDA device is available, falling back to CPU flat scan. For indexed shards (HNSW / IVF-PQ): rayon parallel-map over queries — graph traversal is inherently sequential and has no GPU batch path.

Returns one Vec<SearchResult> per input query, in the same order.

pub fn search_query( &self, query: &[f32], config: &SearchConfig, ) -> Vec<SearchResult>

Search using pre-loaded indexes. No I/O — pure in-memory search.