sqlite_plus_vec 0.1.1

Brute-force vec0 vector search for sqlite_plus, with ORM nearest queries.

sqlite_plus_vec

Nearest neighbors for one float32 column. The module name is vec0. The query reads every stored row and returns the k closest.

dependencies:
  sqlite_plus: ^0.1.7
  sqlite_plus_vec: ^0.1.1

The packages on pub.dev

Each package is a complete library. Import only what you call. The links below are the package pages.

Package	Version	Page
sqlite_plus	^0.1.7	https://pub.dev/packages/sqlite_plus
sqlite_plus_orm	^0.1.7	https://pub.dev/packages/sqlite_plus_orm
sqlite_plus_generator	^0.1.7	https://pub.dev/packages/sqlite_plus_generator
sqlite_plus_arabic	^0.1.1	https://pub.dev/packages/sqlite_plus_arabic
sqlite_plus_vec	^0.1.1	https://pub.dev/packages/sqlite_plus_vec
sqlite_plus_geo	^0.1.0	https://pub.dev/packages/sqlite_plus_geo

sqlite_plus

dependencies:
  sqlite_plus: ^0.1.7

import 'package:sqlite_plus/sqlite_plus.dart';

final db = await openInMemory();
await db.execute(
  'CREATE TABLE note (id INTEGER PRIMARY KEY, title TEXT)',
);
await db.execute('INSERT INTO note(title) VALUES (?)', ['Kept Garden']);
final rows = await db.query(
  'SELECT title FROM note WHERE title LIKE ?',
  ['%Garden%'],
);

rows is Kept Garden. Integers come back as BigInt. There is no FFI and no native SQLite in this engine.

sqlite_plus_orm

dependencies:
  sqlite_plus_orm: ^0.1.7

import 'package:sqlite_plus_orm/sqlite_plus_orm.dart';

@DbEntity()
class Note {
  Note({this.id, required this.title, this.revision = 0});
  final BigInt? id;
  String title;
  @DbVersion()
  int revision;
}

saveChanges writes one transaction and bumps revision. A table with no PRIMARY KEY still maps. Its key is the SQLite row id:

@DbKey(generated: true)
final BigInt? rowid;

plan() adopts that row id and does not rebuild the table.

sqlite_plus_generator

dev_dependencies:
  sqlite_plus_generator: ^0.1.7

dart run sqlite_plus_generator

From the package root, that command writes a part for every public @DbEntity under lib/, plus lib/sqlite_plus_migrations.g.dart.

sqlite_plus_arabic

dependencies:
  sqlite_plus_arabic: ^0.1.1

import 'package:sqlite_plus_arabic/sqlite_plus_arabic.dart';

SqlitePlusArabic.register(db);
await db.execute(
  "CREATE VIRTUAL TABLE passage USING fts5(body, tokenize = 'arabic')",
);

حديقة and الحديقة share one token. الله stays الله. This is not the SQLite ICU Arabic tokenizer. Register before CREATE, and again after the image is reopened.

sqlite_plus_vec

dependencies:
  sqlite_plus_vec: ^0.1.1

import 'package:sqlite_plus_vec/sqlite_plus_vec.dart';

SqlitePlusVec.register(db);
await db.execute(
  'CREATE VIRTUAL TABLE point USING vec0(embedding float[2] distance_metric=l2)',
);

MATCH with k scans every stored float32 vector. The file is not a sqlite-vec database.

sqlite_plus_geo

dependencies:
  sqlite_plus_geo: ^0.1.0

import 'package:sqlite_plus_geo/sqlite_plus_geo.dart';

final geo = GeoSession(db);
await geo.createSpatialTable(
  'facilities',
  geometryColumn: 'location',
  geometryType: GeoGeometryType.point,
  srid: 4326,
);

X is longitude and Y is latitude. ST_SetSRID only labels a geometry. ST_Transform moves a coordinate between EPSG:4326, EPSG:3857, and WGS84 UTM. A degree is not a metre. This is not an OGC, PostGIS, SpatiaLite, or GeoPackage certification.

An application that does not import this package does not load vec0. sqlite_plus stores the rows in a shadow table. This package supplies the distance and the virtual table.

The scan compares every vector. It does not build an approximate index. A file written here is a sqlite_plus image with a shadow table named <table>_vec. A sqlite-vec database file is a different format.

Distances

DbVectorMetrics.l2 is Euclidean distance, the square root of the sum of squared coordinate differences. DbVectorMetrics.cosine is one minus the cosine similarity. Cosine distance sits from 0 to 2. Two zero vectors have distance 0. A zero vector beside a non-zero vector has distance 1. The metric name is compared without regard to case, so 'COSINE' is cosine.

Left	Right	Metric	Distance
[0, 0]	[3, 4]	l2	5
[1, 0]	[1, 0]	cosine	0
[1, 0]	[0, 1]	cosine	1
[1, 0]	[-1, 0]	cosine	2
[0, 0]	[0, 0]	cosine	0
[0, 0]	[1, 0]	cosine	1

vecDistance is that function. Different lengths throw vec0 dimension mismatch. hamming throws vec0: unknown distance metric: hamming.

encodeVec writes little-endian float32. decodeVec reads it back. Dimensions are 1 to 4096, the same bounds as DbVectorLimits. An empty list, a list longer than 4096, and a non-finite coordinate are rejected. decodeVec(encodeVec(const [0.5, -2, 0.25])) is [0.5, -2, 0.25].

SQL

Register before CREATE. SqlitePlusVec.register accepts a SqliteDatabase, a MutableSqliteDatabase, or the connection's SqlFunctionRegistry. A second call keeps the module that is already registered.

import 'package:sqlite_plus/sqlite_plus.dart';
import 'package:sqlite_plus_vec/sqlite_plus_vec.dart';

final db = await openInMemory();
SqlitePlusVec.register(db);
await db.execute(
  'CREATE VIRTUAL TABLE items USING vec0(embedding float[2] distance_metric=l2)',
);
await db.execute(
  'INSERT INTO items(rowid, embedding) VALUES (1, ?), (2, ?), (3, ?)',
  [
    encodeVec(const [1, 0]),
    encodeVec(const [0, 1]),
    encodeVec(const [0.2, 0.1]),
  ],
);
final near = await db.query(
  'SELECT rowid, distance FROM items WHERE embedding MATCH ? AND k = 2',
  [encodeVec(const [0.9, 0.1])],
);

near is row 1, then row 3. Row 1 is [1, 0], the closer of the two to [0.9, 0.1]. Row 2 is [0, 1] and is farther, so k = 2 leaves it out. distance is a double, and the first distance is smaller than the second.

The hidden columns are distance and k. The vector is column 0. MATCH without k throws vec0: k is required. k = 0 throws vec0: k is out of range. In SQL, k is an integer from 1 to 100000. A query with no MATCH returns every row and leaves distance and k null.

Creating the table before register fails with No such module: vec0. distance_metric=hamming fails with vec0: unknown distance metric. float[0] is rejected. A blob whose length is not dimensions * 4 can be inserted into the shadow table, and reading it throws vec0 dimension mismatch.

The bytes live in items_vec. DROP TABLE items drops that shadow. BEGIN, an insert, and ROLLBACK leave count(*) at 0. exportImage keeps the shadow rows. fromImage needs SqlitePlusVec.register again before the next query. The virtual-table name in the schema is not enough. The module object lives on the connection.

A mapped table

sqlite_plus_orm writes the CREATE statement and the nearest method. It does not import this package. DbModules.vec0 is 'vec0'. DbVectorMetrics.l2 is 'l2'. @DbVector(2) uses l2 when metric is omitted. Another identifier can be written as a string. vec0 checks it when the table is created.

import 'package:sqlite_plus_orm/sqlite_plus_orm.dart';
import 'package:sqlite_plus_vec/sqlite_plus_vec.dart';

@DbEntity(module: DbModules.vec0)
class Point {
  Point({this.rowid, required this.embedding});

  @DbKey(generated: true)
  final BigInt? rowid;

  @DbVector(2, metric: DbVectorMetrics.l2)
  final List<double> embedding;
}

The table has one rowid and one List<double>. distance and k are reserved names, so a field with either name is rejected. A vector on an ordinary table or on fts5 is rejected. Dimensions outside 1 to 4096 are rejected. The list element is non-nullable double.

final connection = await MutableDatabaseConnection.openMemory();
SqlitePlusVec.register(connection.database);
final context = DbContext(connection, generatedModel);
await context.migrations.ensureCreated();

final points = context.set(PointMapping.instance);
points.add(Point(embedding: const [1, 0]));
points.add(Point(embedding: const [0, 1]));
await context.saveChanges(); // 2

final nearest = await points.query.nearest(const [0.9, 0.1], k: 1).single();
final again = await context.migrations.plan();

ensureCreated writes CREATE VIRTUAL TABLE "Point" USING vec0("embedding" float[2] distance_metric=l2). nearest.embedding is [1.0, 0.0]. The query text is "embedding" MATCH ? AND "k" = ?, ordered by distance. again is empty.

Codecs.vector(2) checks the list before the write. nearest(const [1], k: 1) throws OrmCodecException because the length is not 2. k: 0 throws RangeError. ORM k stops at DbVectorLimits.maximumNeighbors, which is 4096. The engine's SQL bound of 100000 stays available to a raw query.

A cosine table is the same class with metric: DbVectorMetrics.cosine. The created SQL then contains distance_metric=cosine.

What the generator writes

sqlite_plus_generator reads DbModules.vec0 and @DbVector(2). The part contains virtualModule: "vec0", Codecs.vector(2), vectorDimensions: 2, vectorMetric: "l2", and nearest(List<double> vector, {int k = 10}). The default k in that method is 10. It does not import sqlite_plus_vec.

final point = (await PointQuery(points.query)
        .nearest(const [0.9, 0.1], k: 1)
        .toList())
    .single;

point.embedding[0] is greater than 0.5. That is the vector [1, 0], which is closer to [0.9, 0.1] than [0, 1] is. Register SqlitePlusVec before ensureCreated.

The runnable example

packages/sqlite_plus_vec/example saves [1, 0] and [0, 1], then asks for one neighbor of [0.9, 0.1].

cd packages/sqlite_plus_vec/example
dart run bin/main.dart

1.0 | stable true

1.0 is the first coordinate of the chosen vector, so the row is [1, 0]. stable true means a second plan() is empty.

Where each piece lives

You want	Package	You call
The shadow rows and the SQL engine	sqlite_plus	openInMemory, then MATCH
DbModules.vec0, @DbVector, nearest	sqlite_plus_orm	points.query.nearest
The class turned into a part	sqlite_plus_generator	dart run sqlite_plus_generator
The distance and the vec0 module	this package	SqlitePlusVec.register