English | 中文

Lightning-fast, in-process vector database for Dart & Flutter — powered by Zvec.

Zvec Dart SDK embeds the Zvec vector database in Flutter applications through dart:ffi. It supports vector search, filtered queries, full-text search and local persistent collections.

Features

• Search HNSW, IVF and Flat indexes, filter on scalar fields, and combine vector and full-text results.
• Create typed schemas, insert documents and run synchronous queries through the C API.
• Persist collections across restarts with a write-ahead log and explicit flush.

📦 Installation

flutter pub add zvec

To pin the 0.8 series, use zvec: ^0.8.0 in your pubspec.yaml.

The first Flutter build downloads the target's prebuilt C API library from the official Zvec release and verifies its SHA-256 checksum. Set Android minSdk to 28 or higher, the iOS deployment target to 14.0 or higher, and the macOS deployment target to 11.0 or higher. The macOS framework contains both Intel and Apple Silicon binaries. Platform support · Upstream SDK archives

Upgrading from 0.5.x

• The engine now embeds its default Jieba dictionary and HMM model. Remove calls to ConfigData.setJiebaDictDir and Zvec.setDefaultJiebaDictDir, along with reads of jiebaDictDir and defaultJiebaDictDir. FtsIndexParams(tokenizerName: 'jieba') enables Chinese FTS with the embedded defaults.
• To supply a custom dictionary directory, pass jieba_dict_dir in the JSON string accepted by FtsIndexParams.extraParams. Use user_dict_path to add domain vocabulary. A process-level ZVEC_JIEBA_DICT_DIR also supplies a default directory for fields that have no override. Upstream migration notes
• The GroupByVectorQuery constructor still accepts groupTopk for the maximum number of hits per group.

⚡ Quick Start

import 'dart:typed_data';
import 'package:zvec/zvec.dart';

void main() {
  // 1. Boot the engine
  Zvec.initialize();
  print('Zvec ${Zvec.version}');

  // 2. Define a schema: 4-dim FP32 vector + a string field
  final schema = CollectionSchema(name: 'demo', fields: [
    VectorSchema('embedding', 4, indexParams: HnswIndexParams()),
    FieldSchema(name: 'title', dataType: DataType.string),
  ]);

  // 3. Create & open the collection
  final collection = Collection.createAndOpen('/tmp/zvec_demo', schema);

  // 4. Insert documents
  final docs = [
    Doc(id: 'doc_1')
      ..setField('title', 'hello')
      ..setVector('embedding', Float32List.fromList([0.1, 0.2, 0.3, 0.4])),
    Doc(id: 'doc_2')
      ..setField('title', 'world')
      ..setVector('embedding', Float32List.fromList([0.2, 0.3, 0.4, 0.1])),
  ];
  collection.insert(docs);
  for (final d in docs) d.destroy();

  // 5. Build the index
  collection.optimize();

  // 6. Vector search — top-k by cosine similarity
  final query = VectorQuery(
    fieldName: 'embedding',
    vector: Float32List.fromList([0.4, 0.3, 0.3, 0.1]),
    topk: 5,
    outputFields: ['title'],
  );
  for (final r in collection.query(query)) {
    print('${r.pk}  score=${r.score}  title=${r.getString('title')}');
  }
  query.destroy();

  // 7. Clean up
  collection.close();
  Zvec.shutdown();
}

A full Flutter demo lives in example/lib/main.dart.

Contributing

To run the tests or regenerate bindings, see the developer guide.

📄 License

Released under the Apache License 2.0 — same as upstream Zvec.