Struct EpisodicStore 

pub struct EpisodicStore<'a> { /* private fields */ }

Dedicated persistence view over the episodic-record LMDB database.

Implementations

impl<'a> EpisodicStore<'a>

pub fn with_adaptive_aliases(self, aliases: AdaptiveAliases) -> Self

Attach adaptive aliases for query and ingest-time key expansion.

pub fn with_enrichment(self, enrichment: VocabularyEnrichment) -> Self

Attach vocabulary enrichment for index-time term expansion.

pub fn with_embedder(self, embedder: Arc<dyn Embedder + Send + Sync>) -> Self

Attach an embedder for vector reranking.

pub fn append(&self, record: &EpisodicRecord) -> Result<()>

Append a source record without allowing an existing raw ID to be overwritten.

pub fn append_batch(&self, records: &[EpisodicRecord]) -> Result<()>

Append many source records, then project them into the sidecar in one term-index transaction.

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

Rebuild the DUP_SORT sidecar from the authoritative raw records.

Used when a store contains records but an empty v2 sidecar (legacy stores whose postings lived in the retired v1 database, or stores whose sidecar was lost). The raw lane is never touched; a failure leaves the sidecar empty and search falls back to the raw scan.

pub fn sidecar_is_empty(&self) -> Result<bool>

Number of (term, id) postings in the sidecar.

LMDB 0.8 exposes no per-database stat, so emptiness is checked by peeking the first entry (cheap; whole-db counts are not needed). iter() (not iter_start) is used because iter_start unwraps MDB_FIRST, which fails NotFound on an empty database.

pub fn record_count(&self) -> Result<u64>

Number of raw records in the authoritative lane.

pub fn append_explicit( &self, record: &EpisodicRecord, policy: EpisodicCapturePolicy, ) -> Result<bool>

Append an explicit record according to the capture policy.

pub fn append_explicit_batch( &self, records: &[EpisodicRecord], policy: EpisodicCapturePolicy, ) -> Result<usize>

Append many explicit records according to the capture policy.

pub fn get(&self, id: EpisodicId) -> Result<Option<EpisodicRecord>>

Read an episodic record by its canonical ID.

pub fn transition( &self, id: EpisodicId, transition: MemoryTransition, ) -> Result<()>

Apply an explicit lifecycle transition to a persisted record.

pub fn scan( &self, session_id: Option<Uuid>, limit: usize, ) -> Result<Vec<EpisodicRecord>>

Return records in sequence order, optionally restricted to a session.

pub fn search( &self, query: &str, limit: usize, include_historical: bool, ) -> Result<Vec<EpisodicSearchResult>>

Search current episodic records using deterministic token overlap.

This is a v6 library path only. It does not alter the v5 MCP search route and keeps source records attached to every hit.

When the query asks for the current/latest value of something (see is_current_query), the topic cluster is reordered by deterministic chronology so the most recent statement outranks older ones — post-retrieval temporal resolution, not a scoring change.

pub fn search_with_limits( &self, query: &str, limit: usize, candidate_limit: usize, include_historical: bool, ) -> Result<Vec<EpisodicSearchResult>>

Search with an explicit candidate budget before selective reranking.

Uses multi-query pool widening: the original query plus sub-queries focused on key content words generate candidate IDs. All candidates are then scored with the primary query’s deterministic scoring. This helps answer turns that match few original query terms but match the key entity strongly enter the candidate pool.

pub fn search_with_rerank( &self, query: &str, limit: usize, candidate_limit: usize, include_historical: bool, alpha: f32, ) -> Result<Vec<EpisodicSearchResult>>

Search with vector reranking on top of deterministic scoring.

Pipeline: deterministic scoring → top-N candidates → embed query + candidates → reranking → top-K.

Modes by alpha:

• alpha >= 2.0 — protected top-K (ConvMemory v2 pattern): fully reorder only the deterministic top-limit set by cosine similarity. Membership is fixed, so recall@limit is preserved by construction; only ordering (R@1/MRR) can change.
• 1.0 <= alpha < 2.0 — tiebreaker mode: only reorder adjacent candidates whose deterministic scores are within δ=0.05, using cosine as the tiebreaker.
• alpha < 1.0 — hybrid blending: score = α·det_norm + (1-α)·cosine.

Falls back to search_with_limits() when no embedder is attached.