Struct MessageIndex 

pub struct MessageIndex { /* private fields */ }

sqlite-vec-backed index of chat-message embeddings. Cheap to clone (Arc inside). Stores vectors + metadata only; never message text.

Implementations

impl MessageIndex

pub fn open(path: PathBuf, embedder: Arc<dyn SearchEmbedder>) -> Result<Self>

Open (or create) the message index at path using the supplied embedder. The default db path (~/.ryu/message-embeddings.db) and the registry-driven embedder choice are resolved Core-side by the search_host shim.

pub fn open_in_memory(embedder: Arc<dyn SearchEmbedder>) -> Result<Self>

Open an in-memory index with the supplied embedder at that embedder’s dims. Used by tests (both this crate’s and Core’s, via the search_host shim).

pub fn embedder(&self) -> Arc<dyn SearchEmbedder>

Snapshot the embedder (cheap Arc clone) so embedding I/O never holds a lock.

pub async fn index_message( &self, message_id: &str, conversation_id: &str, role: &str, embedding: &[f32], embed_model: &str, created_at: i64, ) -> Result<()>

Index a single message’s embedding. Idempotent on message_id (re-indexing replaces the prior row + vector). The embedding length must equal the table width; a mismatch is rejected (a vec0 insert would otherwise error).

pub async fn indexed_ids(&self) -> Result<HashSet<String>>

The set of message ids already indexed (used to compute the backfill set).

pub async fn search( &self, query: &str, limit: usize, conversation_ids: Option<&[String]>, ) -> Result<Vec<MessageHit>>

KNN search. Embeds query, runs a cosine-distance KNN over the vec0 table filtered to the current embedder’s model (incomparable vector spaces are skipped), optionally scoping to a set of conversation ids. Returns hits ordered nearest-first.

Trait Implementations

impl Clone for MessageIndex

fn clone(&self) -> MessageIndex

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.