¶Collaborative, syncable vector index

The flagship combination. CRDT offline merge × DiskANN vector search × inline embed() — a shared, semantic knowledge base that lives on the edge and syncs offline, with no central server.

Most vector databases assume a central server. SQL Anywhere ships the three pieces needed to do it the other way around, all in one embedded engine:

• embed() — turn text into a vector inline (no external service)
• DiskANN vector index — approximate nearest-neighbour search over those vectors
• cr-sqlite CRDT — conflict-free, multi-writer offline merge

Put together, several devices can each build a semantic index offline and independently, then merge conflict-free. Afterwards every device can vector-search over every device's documents — the index is maintained as cr-sqlite applies merged rows to the base table.

This is verified: sqlanywhere/tests/collab_vector.rs asserts that a document indexed only on node B becomes the nearest neighbour on node A after a merge.

¶Why this is unusual

¶How

1. Build the cr-sqlite extension (scripts/build-crsqlite.sh) or download it from a release.

2. On each device: load the extension, create a vector-indexed table, and mark it as a conflict-free replicated relation.

   .load ./crsqlite
   CREATE TABLE docs (
     id   INTEGER PRIMARY KEY NOT NULL,
     body TEXT NOT NULL DEFAULT '',
     emb  FLOAT32(128)
   );
   CREATE INDEX docs_vec ON docs(sqlanywhere_vector_idx(emb, 'metric=cosine'));
   SELECT crsql_as_crr('docs');
   

3. Index documents offline, embedding text inline:

   INSERT INTO docs (id, body, emb)
   VALUES (1, 'the cat sat on the mat', vector32(/* embed('the cat sat on the mat') */));
   

   From Rust, embed() does this for you:

   conn.execute(
       "INSERT INTO docs (id, body, emb) VALUES (?, ?, vector32(?))",
       params![id, body, sqlanywhere::embed(body, 128)],
   ).await?;
   

4. Sync by exchanging crsql_changes rows (see docs/CRDT.md). After the merge, search finds documents from all devices:

   SELECT d.body
   FROM   vector_top_k('docs_vec', vector32(/* embed(query) */), 5) k
   JOIN   docs d ON d.id = k.id;
   

¶Runnable example

scripts/build-crsqlite.sh
cargo run -p sqlanywhere --example collab_vector

Before sync — node A only knows its own docs:
  A: search 'vehicle on the road' -> ["the cat sat on the mat", ...]

After sync — each node can search over BOTH devices' documents:
  A: search 'vehicle on the road' -> ["the car drove down the highway",
                                       "a truck delivered the heavy cargo", ...]

Source: sqlanywhere/examples/collab_vector.rs.

¶Notes & tips

• The bundled embed() is a lexical reference embedder — great for a zero-dependency demo. For real semantic quality, compute embeddings with a model and store the numbers the same way; the CRDT + index behaviour is identical.
• Use compress_neighbors to keep the synced index small on constrained devices.
• Combine with hybrid search (FTS5 + vector) for production-grade retrieval over the merged corpus.
• Status: experimental, tracking the cr-sqlite extension (not yet bundled in sqld).