Search

Search

Every any server keeps a local search index — BM25 full-text plus semantic vectors — over the chats, documents, property values and runtime-dataset records of every space it holds. One endpoint, POST /v1/spaces/:spaceId/search, queries it in three modes; nothing leaves the machine.

The shape of it

   space change feed            index.db (per space)             /search
 ┌──────────────────┐   chunkers   ┌───────────────────┐  fts    ┌──────────┐
 │ chat_messages    │ ───────────▶ │ doc: objectId:    │ ──────▶ │          │
 │ editor_blocks    │  text only   │      dataset:     │         │  hybrid  │
 │ objects (props)  │              │      recordId     │  vector │  (RRF)   │
 │ runtime datasets │              │ data, scope, hash │ ──────▶ │          │
 └──────────────────┘              │ vector?, pending? │         └──────────┘
                                   └───────────────────┘

• Indexing is a background consumer of each space's change feed. It never blocks writes and never touches the CRDT — the index is derived state that can be deleted and rebuilt.
• Full-text search is always available, with no external dependency. New writes are searchable within about a 250 ms debounce.
• Vector search activates when an embedder is configured. The default runs an embedding model in-process (llama.cpp, auto-downloaded); an unreachable embedder only pauses the vector side — full-text keeps working.
• Hybrid is the default query mode: both legs fused by reciprocal rank, degrading to full-text on its own when the embedder cannot help. The reply says which mode actually ran.

Why it matters. Search over an encrypted, local-first database has to run where the plaintext is — on the device. There is no server-side index to leak, no query log anywhere else, and the same index works offline.

A first query

curl -s http://127.0.0.1:7001/v1/spaces/$SPACE/search \
  -H 'content-type: application/json' \
  -d '{"query": "zeppelin disaster", "limit": 5}'

any search $SPACE "zeppelin disaster" --limit 5

{
  "hits": [
    { "scope": "chat", "objectId": "…", "dataset": "chat_messages",
      "recordId": "…", "data": "the zeppelin disaster of 1937",
      "score": 0.0328 }
  ],
  "mode": "hybrid",
  "vectorStatus": "used"
}

Hits carry identity, not full records — hydrate them with a dataset query when you need the whole row.

What is indexed

Programs, miniapps and file bytes are never indexed.

Reading further