DB
Back homeIndustry

AI & ML platforms

Retrieval, online features, and semantic search over live data — on one engine that stores vectors next to the rows they describe.

The problem

Why this is hard today

AI platforms accrete stores: a vector database for embeddings, a feature store for online features, a warehouse for training data, and the system of record underneath. Every boundary is a sync job, and every sync job is a freshness gap the model pays for at inference time.

The gaps compound: retrieval answers from a stale corpus, features are read from a snapshot that predates the current request, and semantic search indexes lag the data they describe.

A single engine with first-class vector(N) columns, HNSW ANN in SQL, keyed online features with read-your-writes, and continuous transforms to keep it all fresh removes the sync layer instead of scheduling it.

Where NYXDB fits

Use-case journeys

RAG retrieval infrastructure

Store embeddings as vector(N), run ANN search that rewrites to VectorTopK, and keep context fresh with transforms.

AI & vector search

Online feature serving

Keyed tables serve online features with read-your-writes and exact counts — no separate feature store.

Fraud & risk

Semantic search over live data

Index embeddings that stay fresh as the underlying data changes, driven by continuous transforms.

AI & vector search

Architecture

How NYXDB fits AI & ML platforms

Embeddings, source rows, and online features live in one engine; HNSW indexes power ANN search; transforms keep retrieval and feature tables fresh from the stream.

  1. 01

    Embed

    vector(N) columns store embeddings next to source data.

  2. 02

    Index

    Per-part HNSW (usearch) built at flush; cosine default.

  3. 03

    Serve

    VectorTopK ANN search and keyed online features in one SQL surface.

  4. 04

    Refresh

    Transforms materialize fresh embeddings and features from the stream.

Real SQL

Representative query

ANN search over embeddings
SELECT id, cosine_distance(emb, '[1,0,0,0,0,0,0,0]')
FROM emb_t
ORDER BY cosine_distance(emb, '[1,0,0,0,0,0,0,0]')
LIMIT 3;

Every statement follows the engine’s own test SQL shapes. See the SQL reference for full syntax.

Capabilities

What you get

First-class vectors

vector(N) columns next to source rows.

HNSW ANN

Per-part index; ORDER BY distance → VectorTopK.

Online features

Keyed tables, read-your-writes.

Always fresh

Transforms keep retrieval + features live.

Proof

Measured where it counts

~14–16×

vector(N) kernel speedup

PR #370 / #131 (kernel)

VectorTopK

ANN rewrite (AnnRewriteRule)

ADR-064

end-to-end ANNTODO-verify

query latency class

placeholder — not yet substantiated

Figures marked TODO-verify are placeholders pending a published, reproducible benchmark; substantiated numbers cite their source.

One engine for the AI data layer

Explore AI & vector search or run a node.