Struct SemanticIndex 

pub struct SemanticIndex { /* private fields */ }

Three-index semantic search over the stage store.

Implementations

impl SemanticIndex

pub fn from_stages( stages: Vec<Stage>, provider: Box<dyn EmbeddingProvider>, config: IndexConfig, ) -> Result<Self, EmbeddingError>

Build the index from an owned list of stages (useful in async contexts where holding a &dyn StageStore across .await is not possible).

pub fn build( store: &dyn StageStore, provider: Box<dyn EmbeddingProvider>, config: IndexConfig, ) -> Result<Self, EmbeddingError>

Build the index from all non-tombstoned stages in a store.

pub fn from_stages_batched( stages: Vec<Stage>, cached_provider: CachedEmbeddingProvider, config: IndexConfig, chunk_size: usize, ) -> Result<Self, EmbeddingError>

Build the index in a single pass: collect every signature/description/ example text upfront, dispatch all cache misses through inner.embed_batch in chunks of chunk_size, then assemble the three sub-indexes. Used by noether-cloud’s registry on cold start so that 486 stages × 3 texts = 1458 individual API calls collapse into ~46 batch calls of 32 texts each — well within typical rate limits.

pub fn from_stages_batched_paced( stages: Vec<Stage>, cached_provider: CachedEmbeddingProvider, config: IndexConfig, chunk_size: usize, inter_batch_delay: Duration, ) -> Result<Self, EmbeddingError>

Like from_stages_batched, but waits inter_batch_delay between successive batch calls and commits cache entries to disk after each batch. Use this with rate-limited remote providers (e.g. Mistral free tier ≈ 1 req/s → pass ~1100 ms).

pub fn build_cached( store: &dyn StageStore, cached_provider: CachedEmbeddingProvider, config: IndexConfig, ) -> Result<Self, EmbeddingError>

Build using a CachedEmbeddingProvider for persistent embedding cache.

pub fn add_stage(&mut self, stage: &Stage) -> Result<(), EmbeddingError>

Add a single stage to all three indexes.

pub fn remove_stage(&mut self, stage_id: &StageId)

Remove a stage from all three indexes.

pub fn len(&self) -> usize

Number of stages indexed.

pub fn is_empty(&self) -> bool

pub fn search( &self, query: &str, top_k: usize, ) -> Result<Vec<SearchResult>, EmbeddingError>

Search across all three indexes and return ranked results.

pub fn search_filtered( &self, query: &str, top_k: usize, tag: Option<&str>, ) -> Result<Vec<SearchResult>, EmbeddingError>

Like search, but restricts candidates to stages carrying tag (exact match). Passing tag: None is equivalent to search.

pub fn search_by_tag(&self, tag: &str) -> Vec<StageId>

Return all stage IDs that carry tag (exact match).

pub fn all_tags(&self) -> Vec<String>

Return the set of all known tags across indexed stages.

pub fn check_duplicate_before_insert( &self, description: &str, threshold: f32, ) -> Result<Option<(StageId, f32)>, EmbeddingError>

Check whether a candidate description is a near-duplicate of an existing stage.

Returns Some((stage_id, similarity)) if any existing stage’s semantic embedding exceeds threshold (default 0.92). Returns None if the description is novel enough.

pub fn find_near_duplicates( &self, threshold: f32, ) -> Vec<(StageId, StageId, f32)>

Scan all active stages for near-duplicate pairs.

Returns pairs (id_a, id_b, similarity) where semantic similarity >= threshold. Each pair appears only once (id_a < id_b lexicographically).