Struct RecallEngine 

pub struct RecallEngine { /* private fields */ }

Hybrid recall engine combining BM25 + vector search.

Wraps a MemoryStore, SearchEngine, VectorStore, and Embedder to provide fused search at recall time and auto-embedding at write time.

Implementations

impl RecallEngine

pub fn new( store: Arc<MemoryStore>, search_engine: Arc<SearchEngine>, vector_store: VectorStore, embedder: Arc<dyn Embedder>, config: RecallConfig, ) -> Result<Self>

Create a new recall engine.

Returns an error if the Tantivy IndexWriter cannot be created.

pub fn record_relevance_feedback( &self, score: f32, relevant: bool, ) -> Result<usize>

Record one relevance-feedback sample into the conformal calibrator (V8 S8). Errors honestly when the knob is off — there is no calibrated set to feed.

pub fn conformal_disclosure( &self, results: &mut [RecallResult], ) -> Result<Option<ConformalSetInfo>>

Conformal disclosure for a completed search: grades the results against the calibrated set (marking in_conformal_set) and returns the set-level info. Ok(None) when the knob is off — no claim is made at all.

pub fn config(&self) -> &RecallConfig

Get the configuration.

pub fn embedder_is_real(&self) -> bool

Whether the embedder is a real backend (not a stub).

When false, hybrid search would produce garbage vectors and store_with_embedding should be avoided in favor of plain BM25.

pub fn embedder_probe(&self) -> Result<usize>

Probe the configured embedder with one tiny input.

Degradation honesty: a configured embedder that cannot answer (server down, model missing) must not silently downgrade recall. Returns the produced vector length on success.

Errors

Propagates the embedder’s error (transport, model, dimension).

pub fn store_with_embedding( &self, galaxy: Galaxy, memory: &Memory, ) -> Result<()>

Store a memory with auto-embedding.

1. Embeds the content using the configured embedder.
2. Stores the memory in the given galaxy.
3. Stores the embedding in the Embeddings galaxy.
4. Adds the embedding to the vector store.
5. Indexes the content in Tantivy.

pub fn store_batch_with_embedding( &self, entries: &[(Galaxy, &Memory)], ) -> Result<usize>

Batch-store memories with auto-embedding in a single HTTP call + single Tantivy commit.

Like store_with_embedding but for multiple memories at once:

1. Embeds all content via embed_batch() (single HTTP call).
2. Stores all memories to LMDB.
3. Stores all embeddings.
4. Adds all to the vector store.
5. Indexes all in Tantivy with a single writer + commit.

Returns the number of memories successfully stored.

pub fn backfill_embeddings( &self, galaxy: Option<Galaxy>, limit: usize, dry_run: bool, ) -> Result<BackfillReport>

Backfill per-memory vectors for memories that have none.

Streams the requested galaxies in LMDB key order and collects up to limit memories lacking a stored embedding (limit == 0 = no cap), then embeds + persists them, also seeding the in-memory vector index. The scan opens its own read transaction and the embed/put writes use their own transactions — no nested LMDB read txns, no full-galaxy materialization (candidate collection is capped by limit). dry_run reports candidates without writing.

Errors

Fails fast when the wired embedder is the stub (backfill would store noise), matching the embedder_is_real gate used by the write path.

pub fn hybrid_search( &self, query: &str, limit: usize, galaxy_filter: Option<Galaxy>, ) -> Vec<RecallResult>

Hybrid search combining BM25 + vector similarity.

Weights: bm25_weight * BM25 + vector_weight * cosine + importance_weight * importance

pub fn hybrid_search_with_disclosure( &self, query: &str, limit: usize, galaxy_filter: Option<Galaxy>, ) -> (Vec<RecallResult>, Option<ConformalSetInfo>)

Hybrid search plus the V8 S8 disclosure: (results, conformal). conformal is None when WM_RECALL_CONFORMAL_ALPHA is unset — no calibrated claim exists, so none is made.

pub fn vector_search( &self, query: &str, limit: usize, galaxy_filter: Option<Galaxy>, ) -> Vec<RecallResult>

Pure vector search (no BM25).

pub fn text_search(&self, query: &str, limit: usize) -> Vec<RecallResult>

Pure BM25 search (no vector).

pub fn promote_memory(&self, galaxy: Galaxy, id: Uuid) -> Result<bool>

Promote a memory on recall hit: calls Memory::recall() to apply Hebbian strengthening and updates accessed_at/access_count/recall_count in the store.

pub fn is_private(&self, id: Uuid, galaxy: Galaxy) -> bool

Whether a memory is flagged is_private (missing memories count as private — they cannot be verified visible).

pub fn cache_size(&self) -> usize

Get the number of cached embeddings.

pub fn clear_cache(&self)

Clear the embedding cache.

pub fn vector_count(&self) -> usize

Get the number of vectors in the vector store.