DB
Back homeUse case

AI & vector search

Store embeddings as a first-class column, run ANN search in SQL, and keep retrieval context fresh from live streams — no separate vector database.

The problem

Why this is hard today

RAG and semantic-search stacks bolt a dedicated vector database onto the side of the system of record. Embeddings live in one store, the source rows in another, and a batch job shuttles between them — so retrieval is always answering from a stale snapshot of the corpus.

The freshness gap is the whole problem for retrieval-augmented generation: an assistant that retrieves last night’s context cannot answer about what happened this morning. And when the online features that drive a model live in yet another store, read-your-writes is lost exactly where inference needs it.

What AI platforms need is embeddings stored next to the data, approximate-nearest-neighbor search expressed in the same SQL as everything else, and a way to keep the vector index fresh from the event stream instead of a nightly re-embed.

Architecture

How NYXDB does it

Embeddings are a vector(N) column; a per-part HNSW index is built at flush; an ANN query rewrites to the VectorTopK operator. Continuous transforms keep embedding-enriched tables fresh from the stream.

  1. 01

    Store

    Embeddings live in a vector(N) column alongside the source columns — one table, one row.

  2. 02

    Index

    A per-part HNSW index (vendored usearch) is built at flush; bare INDEX hnsw defaults to the cosine metric.

  3. 03

    Search

    ORDER BY cosine_distance(emb, $q) LIMIT k rewrites to the VectorTopK ANN operator (AnnRewriteRule).

  4. 04

    Stay fresh

    A CREATE TRANSFORM materializes embedding-enriched rows from the event stream — fresh retrieval context, no batch re-embed.

Real SQL

In practice

Embeddings with an HNSW index
CREATE TABLE emb_t (
  id  UInt64,
  ATTRIBUTE (emb vector(8) INDEX hnsw)
);
ANN search — rewrites to VectorTopK
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;
Streaming RAG — keep context fresh
-- materialize embedding-enriched rows from the event stream
CREATE TRANSFORM enrich_docs
  INTO doc_embeddings AS
SELECT doc_id, emb
FROM incoming_docs;

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

Capabilities

What you get

First-class embeddings

vector(N) columns store embeddings next to source data — no separate store.

HNSW ANN

Per-part HNSW index (usearch) built at flush; ORDER BY distance → VectorTopK.

Cosine + more

cosine, L2, and inner-product distance functions; bare INDEX hnsw defaults to cosine.

Streaming RAG

CREATE TRANSFORM keeps embedding-enriched tables fresh from the stream.

Online features

Keyed tables serve online features with read-your-writes and exact counts.

One engine

Vectors, source rows, and features share one runtime and one SQL surface.

Proof

Measured where it counts

~14–16×

vector(N) raw-ABI kernel speedup

PR #370 / #131 (kernel, not end-to-end)

VectorTopK

ANN rewrite over the HNSW index

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.

Learn more

Related documentation

Vector search

vector(N), HNSW, VectorTopK, cosine.

Data types

The vector(N) type.

Transforms

Materialize fresh embeddings.

Retrieval that is always fresh

Store embeddings, run ANN search, and keep context live — all in SQL.

Get started Read the docs