MoteDB: Embedded multi-modal database for edge AI

#ai #database #rust #showdev

Hi everyone,

I’ve been working on something for the past few months and finally feel it’s ready to show. It’s called MoteDB – an embedded database that’s trying to be a good fit for the kind of AI stuff happening on edge devices right now.

The short version: It’s an embedded, multi-model database (think vectors, text, spatial, time-series, plain old columns) written in Rust, designed to run inside robots, AR glasses, or industrial gizmos without needing a separate server process.

Why bother? When I was playing with some hobby robotics stuff, I kept running into the same friction: I needed to store sensor readings (time-series), find similar camera frames (vectors), remember object locations (spatial), and keep some config data (columns) – all in one place. Bolt on a vector DB here, a time-series DB there, maybe SQLite for the rest… it gets messy fast. I wanted something that just treats all these as native citizens from the start.

So MoteDB gives you:

A SQL interface (yes, real SQL – subqueries, joins, the works) for the boring stuff.
Native multi-modal indexes: You can throw vectors, text, or coordinates at it and query them directly. The index types (DiskANN-like for vectors, BM25 for text, R-tree for spatial) are built-in, not afterthoughts.
Embedded: It links into your Rust app. No network, no container, no drama.
MVCC transactions with savepoints, because even robots need consistent state.
It’s written in Rust. (Obviously, given where I’m posting this.) The whole thing is no_std friendly in core, with the standard runtime for the SQL parser and such.

A quick taste:

// Create a table
db.execute("CREATE TABLE items (id INT, embedding VECTOR(128), tag TEXT, location POINT)")?;

// Insert some multi-modal data
db.execute("INSERT INTO items VALUES (1, '[0.1, 0.2, ...]', 'camera_frame', POINT(10.5, 20.3))")?;

// Find similar vectors (KNN) within a spatial range
let results = db.query(
    "SELECT id, vdistance(embedding, '[0.15, 0.25, ...]') as score 
     FROM items 
     WHERE MBRContains(location, ST_GeomFromText('POLYGON((...))'))
     ORDER BY score 
     LIMIT 5"
)?;

Performance-wise, it’s not messing around:

Batch inserts: ~700k rows/sec on my test hardware.
Vector search: <5ms latency at 95% recall.
Memory footprint: Core structures sit under 100MB.

It’s still early days (v0.1.0 just dropped), but the core is solid and I’m actively using it in a couple of projects. The API is stabilizing, and I’d love for more folks to kick the tires.

If you’re into: