Struct EmbeddingEngine 

pub struct EmbeddingEngine { /* private fields */ }

Implementations

impl EmbeddingEngine

pub fn find_similar_code( &self, file_path: &str, symbol_name: &str, max_results: usize, ) -> Result<Vec<SemanticMatch>>

Find code symbols most similar to the given symbol.

pub fn find_duplicates( &self, threshold: f64, max_pairs: usize, ) -> Result<Vec<DuplicatePair>>

Find near-duplicate code pairs across the codebase. Returns pairs with cosine similarity above the threshold (default 0.85).

pub fn find_duplicates_in_scope( &self, threshold: f64, max_pairs: usize, scope: Option<&str>, ) -> Result<Vec<DuplicatePair>>

Find near-duplicate code pairs, using scoped anchors when scope is provided.

Candidate search remains global, so cross-boundary duplicates remain visible without paying a full-corpus anchor scan for narrow scopes.

impl EmbeddingEngine

pub fn classify_symbol( &self, file_path: &str, symbol_name: &str, categories: &[&str], ) -> Result<Vec<CategoryScore>>

Classify a code symbol into one of the given categories using zero-shot embedding similarity.

pub fn find_misplaced_code( &self, max_results: usize, ) -> Result<Vec<OutlierSymbol>>

Find symbols that are outliers — semantically distant from their file’s other symbols.

pub fn embed_text(&self, text: &str) -> Result<Vec<f32>>

pub fn store_artifact_embeddings( &self, chunks: &[ArtifactEmbeddingChunk], ) -> Result<usize>

Store pre-computed artifact embeddings in the semantic index.

pub fn search_artifact_embeddings( &self, query: &str, top_k: usize, ) -> Result<Vec<ScoredArtifactChunk>>

Semantic search over stored artifact analyses.

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

Count stored artifact embeddings.

pub fn prune_artifact_embeddings(&self, max_age_ms: u64) -> Result<usize>

Prune artifact embeddings older than the given duration (ms).

pub fn file_mean_embeddings( &self, file_paths: &[&str], ) -> Result<HashMap<String, Vec<f32>>>

Compute mean embedding for each file from indexed symbol embeddings.

pub fn mean_of_embeddings(embeddings: &[Vec<f32>]) -> Option<Vec<f32>>

Compute mean embedding of multiple file embeddings.

impl EmbeddingEngine

pub fn new(project: &ProjectRoot) -> Result<Self>

pub fn model_name(&self) -> &str

pub fn runtime_info(&self) -> &EmbeddingRuntimeInfo

pub fn is_indexing(&self) -> bool

Returns true if a full reindex is currently in progress.

pub fn index_from_project(&self, project: &ProjectRoot) -> Result<usize>

pub fn ensure_index_fresh_for_project( &self, project: &ProjectRoot, ) -> Result<EmbeddingFreshnessReport>

pub fn generate_bridge_candidates( &self, project: &ProjectRoot, ) -> Result<Vec<(String, String)>>

Extract NL→code bridge candidates from indexed symbols. For each symbol with a docstring, produces a (docstring_first_line, symbol_name) pair. The caller writes these to .codelens/bridges.json for project-specific NL bridging.

pub fn index_changed_files( &self, project: &ProjectRoot, changed_files: &[&str], ) -> Result<usize>

Incrementally re-index only the given files.

pub fn is_indexed(&self) -> bool

Whether the embedding index has been populated.

pub fn index_info(&self) -> EmbeddingIndexInfo

pub fn inspect_existing_index( project: &ProjectRoot, ) -> Result<Option<EmbeddingIndexInfo>>

impl EmbeddingEngine

pub fn embed_query_cached(&self, query: &str) -> Result<Vec<f32>>

pub fn prewarm_queries(&self, queries: &[String]) -> Result<usize>

pub fn query_cache_stats(&self) -> Result<QueryEmbeddingCacheStats>

impl EmbeddingEngine

pub fn search( &self, query: &str, max_results: usize, ) -> Result<Vec<SemanticMatch>>

Search for symbols semantically similar to the query.

pub fn search_scored( &self, query: &str, max_results: usize, ) -> Result<Vec<ScoredChunk>>

Search returning raw ScoredChunks with optional reranking.

Pipeline: bi-encoder → candidate pool (3× requested) → rerank → top-N. Reranking uses query-document text overlap scoring to refine bi-encoder cosine similarity. This catches cases where embedding similarity is high but the actual text relevance is low (or vice versa).

pub fn search_scored_in_scope( &self, query: &str, max_results: usize, path_scope: Option<&str>, ) -> Result<Vec<ScoredChunk>>

Search raw chunks within an optional relative path scope.

Scoped searches push the broad path partition and exact/prefix file constraints into sqlite-vec before KNN candidate selection. This keeps scoped ranked-context calls from spending candidate budget on unrelated packages or generated artifacts.