Module store 

Embedding storage: read/write embeddings from/to the structural index SQLite.

Operates through direct libsql calls on the same connection as FileIndex. The embeddings table is created lazily on first write.

When sqlite-vec is available (vec_embeddings virtual table exists), inserts and deletes are mirrored there so that ann_search can use the ANN index instead of loading all vectors into memory.

Constants

ANN_CANDIDATE_COUNT

Number of ANN candidates to retrieve before staleness re-ranking.

Functions

ann_search

ANN search using the vec_embeddings virtual table.

count_embeddings

Count embeddings stored for a given model.

delete_embeddings_for_path

Delete embeddings for a specific (source_type, source_path) pair, all models. Used during incremental rebuild when a file changes.

drop_embedding_tables

Drop embedding tables entirely for a full rebuild.

embedded_paths

Return the set of file paths that have at least one embedding for the given model.

ensure_schema

Ensure the embeddings table and indices exist.

ensure_vec_schema

Ensure the vec_embeddings ANN virtual table exists.

load_all_embeddings

Load all stored embeddings for a given model name.

load_embeddings_for_type

Load stored embeddings filtered by a specific source type.

upsert_embedding

Insert or replace one embedding row.

vacuum

Run VACUUM to reclaim space after a full rebuild.

vec_table_available

Returns true if the vec_embeddings virtual table exists and is queryable.