Struct Bm25Cache 

pub struct Bm25Cache { /* private fields */ }

A lazily-built, shared Bm25Index over a store’s whole chunk corpus.

Building the index reads every chunk, so a retriever that rebuilt it per query paid that cost several times for a single question — hybrid runs one keyword search per query, multi-query one per rewrite. The cache builds it once and hands out an Arc.

Freshness is checked on every use against a cheap store fingerprint (the document and chunk counts, one COUNT(*) each) plus a generation counter that Bm25Cache::invalidate bumps — the pipeline calls that on every write, which is what catches an edit that happens to leave the counts unchanged. A different process writing to the same database is only caught by the counts, so a shared store can serve one stale result after such a same-count edit.

Implementations

impl Bm25Cache

pub fn new() -> Self

An empty cache.

pub fn invalidate(&self)

Drop the cached index. Call after any write to the store.

pub async fn index( &self, store: &Arc<dyn VectorStore>, params: Bm25Params, ) -> Result<Arc<Bm25Index>>

The index for store, rebuilding it if the corpus or params moved.

Trait Implementations

impl Default for Bm25Cache

fn default() -> Self

Returns the “default value” for a type.