ryu-search

Conversation search primitive for Ryu. Part of Ryu.

ryu-search indexes past chat messages for recall. It pairs a sqlite-vec vec0 semantic KNN index (MessageIndex) with a contentless FTS5 lexical index (MessageFtsIndex). Both stores hold vectors / inverted-index + metadata only — never message text; the caller re-reads and decrypts each hit's snippet from conversations.db.

Role in the decomposition

An extracted Core capability crate compiled in-process as a non-optional path dependency — the ConversationStore reaches it unconditionally. ZERO dependency on apps/core (it never sees ModelRegistry).

Swap-seam: the narrow SearchEmbedder trait (dims / model_id / is_local / embed). Core wraps its registry-configured retrieval::Embedder behind this trait object at construction, so which embedder (local hashing vs. remote /v1/embeddings) is a per-consumer RAG choice that stays out of the crate. The default db paths (~/.ryu/message-embeddings.db, ~/.ryu/message-fts.db) and the embedder choice live in the Core host shim apps/core/src/search_host.rs, mirroring the ryu-storage open(path) precedent.

What it provides

• MessageIndex (message_index.rs) — semantic vec0 KNN index → MessageHit.
• MessageFtsIndex (message_fts.rs) — contentless FTS5 lexical index → MessageFtsHit.
• SearchEmbedder — the injection seam, plus a bundled LocalHashingEmbedder default.

Placement (CLAUDE.md §1): searching Core's own conversation history is what runs, so Core.

Build

cargo build -p ryu-search
cargo test  -p ryu-search

License

Apache-2.0; see LICENSE. © 2026 A Major Pte. Ltd.