mediapipe_text

Google's MediaPipe text tasks for Dart and Flutter: text classification, text embeddings (EmbeddingGemma included) and language detection on every platform, plus Proofreader and Summarizer on every platform but the web. Every task runs Google's official MediaPipe pipeline, and Google's pinned models are bundled with your app at build time.

Tasks and platforms

Task	Class	Android	iOS	macOS arm64	Linux x64	Windows x64	Web
Text classification (BERT)	TextClassifier	✓	✓	✓	✓	✓	✓
Text embeddings (Universal Sentence Encoder)	TextEmbedder	✓	✓	✓	✓	✓	✓
Language detection	LanguageDetector	✓	✓	✓	✓	✓	✓
EmbeddingGemma 300M	TextEmbedder	✓	✓	✓	✓	✓	✓
Proofreader 200M	TextProofreader	✓	✓	✓	✓	✓
Summarizer 200M	TextSummarizer	✓	✓	✓	✓	✓

All text tasks run on the CPU. The runtimes are Google's: its MediaPipe 1.2.0 text library on Android 9+, iOS 15+ (devices and the arm64 simulator), macOS 14+, Linux and Windows, which this package's hook bundles, and @mediapipe/tasks-text 1.2.0-rc.20261018 in browsers.

Google's browser runtime has no Proofreader or Summarizer (UP-034), which their capability queries report there.

No platform needs a setting. See platform setup for permissions, Linux graphics libraries and self-hosting the browser runtime.

Quick start

Add mediapipe_text: ^0.1.2 to your app's dependencies, then:

import 'package:mediapipe_text/mediapipe_text.dart';

Future<void> classify() async {
  final classifier = await TextClassifier.create(
    TextClassifierOptions(model: TextModels.bertClassifier),
  );
  try {
    final result = await classifier.classify('Hello, world!');
    final top = result.classifications.first.categories.first;
    print('${top.categoryName}: ${top.score}');
  } finally {
    await classifier.dispose();
  }
}

TextModels names Google's model for every task (bertClassifier, universalSentenceEncoder, languageDetector, embeddingGemma, proofreader, summarizer). Bundle the ones your app uses at build time: list them in its pubspec, declare the folder they go in, and run dart run mediapipe_core:bundle_models from the app's root, which downloads each once and checks it against its SHA-256:

flutter:
  assets:
    - assets/mediapipe/

hooks:
  user_defines:
    mediapipe_text:
      models: [bert_classifier]

TextModels.byName lists the accepted names. A model that is not bundled makes create throw a RuntimeUnavailableException naming the entry to add; setting ModelStore.allowDownloads = true from mediapipe_core downloads it at run time instead. The text generation models are 118 to 184 MB, so weigh that against your app's size. To use your own model, pass modelPath (a file on native platforms, a URL in browsers) or modelBytes (for example from rootBundle); the Proofreader and Summarizer read their large models from a file, so they take model or modelPath. Give exactly one model source.

Results own their data and stay valid after the task is disposed.

Language detection

LanguageDetector returns language codes with probabilities, most likely first.

import 'package:mediapipe_text/mediapipe_text.dart';

Future<void> detectLanguage(String text) async {
  final detector = await LanguageDetector.create(
    LanguageDetectorOptions(model: TextModels.languageDetector),
  );
  try {
    final result = await detector.detect(text);
    for (final prediction in result.predictions.take(3)) {
      print('${prediction.languageCode}: ${prediction.probability}');
    }
  } finally {
    await detector.dispose();
  }
}

Text embeddings

TextEmbedder turns a sentence into a vector with the Universal Sentence Encoder; the static TextEmbedder.cosineSimilarity compares two, including quantized ones, computed in Dart so every platform gives the same answer. The higher the value, the closer the meaning.

import 'package:mediapipe_text/mediapipe_text.dart';

Future<double> compareSentences(String first, String second) async {
  final embedder = await TextEmbedder.create(
    TextEmbedderOptions(model: TextModels.universalSentenceEncoder),
  );
  try {
    final a = await embedder.embed(first);
    final b = await embedder.embed(second);
    return TextEmbedder.cosineSimilarity(
      a.embeddings.first,
      b.embeddings.first,
    );
  } finally {
    await embedder.dispose();
  }
}

Behavior

• create reports initialization errors and refuses a delegate the capability query rules out; each task runs on its own isolate and queues concurrent calls. Every task has a delegate getter.
• dispose() drains accepted calls, releases the native task and waits for the isolate to exit. It is idempotent; calls after it throw StateError.
• Failures are MediaPipeExceptions: RuntimeUnavailableException (the platform or settings cannot run the task, with a fix), ModelDownloadException, and TaskException with Google's message and native statusCode. Errors arrive through the returned Future or stream.
• Input containing NUL characters is rejected before it reaches native code.
• Nothing is reimplemented in Dart: tokenization, prompts, label order, thresholds, normalization and quantization are Google's.

Query support before offering a task or delegate:

import 'package:mediapipe_text/mediapipe_text.dart';

Future<bool> canSummarize() async {
  final support = await queryTextSummarizerCapabilities();
  return support.supportedDelegates.contains(Delegate.cpu);
}

The result also carries the reason for every unavailable delegate. It describes declared support; it does not load a model.

EmbeddingGemma 300M

TextEmbedder with TextModels.embeddingGemma runs Google's complete EmbeddingGemma pipeline: prompt formatting, SentencePiece tokenization, inference and postprocessing. It returns 768-value vectors on every platform, with Google's prompt formatting on each runtime; queryTextEmbedderCapabilities(TextModels.embeddingGemma) reports it.

import 'package:mediapipe_text/mediapipe_text.dart';

Future<double> similarity(String a, String b) async {
  final task = await TextEmbedder.create(
    TextEmbedderOptions(model: TextModels.embeddingGemma),
  );
  try {
    final context = TextFormatContext(
      taskType: EmbeddingType.semanticSimilarity,
    );
    final first = await task.embed(a, formatContext: context);
    final second = await task.embed(b, formatContext: context);
    return TextEmbedder.cosineSimilarity(
      first.embeddings.single,
      second.embeddings.single,
    );
  } finally {
    await task.dispose();
  }
}

All eight official formatting modes are supported, including retrieval documents with titles and query/document roles; omitting formatContext passes none to Google's API. l2Normalize and quantize (both off by default) go to Google's postprocessor, and quantized results keep Google's signed-int8 bytes. The model's 512-token limit includes formatting and special tokens; input is never truncated or chunked. The model is 183.8 MB and under the Gemma Terms.

Proofreader 200M

TextProofreader runs Google's pipeline, including tokenization, generation and the ordered corrections from Google's diff.

import 'package:mediapipe_text/mediapipe_text.dart';

Future<void> proofread() async {
  final task = await TextProofreader.create(
    TextProofreaderOptions(model: TextModels.proofreader),
  );
  try {
    final result = await task.proofread('She go home.');
    print(result.proofreadText);
    for (final edit in result.corrections) {
      print('${edit.type.name}: ${edit.text}');
    }
    await for (final update in task.proofreadStream('I recieved your mesage.')) {
      if (update.chunk != null) print(update.chunk);
      if (update.done) print(update.corrections);
    }
  } finally {
    await task.dispose();
  }
}

Chunks contain newly generated text; the final update carries the complete corrections, each same, insertion or deletion in native order.

Summarizer 200M

TextSummarizer runs Google's pipeline in either official mode, TextSummarizerMode.keypoints (the default) or .tldr, fixed at creation.

import 'package:mediapipe_text/mediapipe_text.dart';

Future<void> summarize(String article) async {
  final task = await TextSummarizer.create(
    TextSummarizerOptions(
      model: TextModels.summarizer,
      mode: TextSummarizerMode.tldr,
    ),
  );
  try {
    print((await task.summarize(article)).summary);
    await for (final update in task.summarizeStream(article)) {
      if (update.chunk != null) print(update.chunk);
    }
  } finally {
    await task.dispose();
  }
}

Google's output is kept as is, bullet markers and whitespace included.

Streaming, token budgets and cancellation

For Proofreader and Summarizer, streams start on listen and allow one subscription; pausing buffers updates. Google's API has no cancellation, so cancelling a stream stops delivery but waits for the native call to finish. maxNumTokens and cacheDirectory go to Google's options (null or zero selects the native default). Budgets include input and output, so an input that already exceeds the budget returns a native error rather than being truncated. Both models are 117.6 MB and under the Gemma Terms.

Validation

Every task is compared with Google's own output for the same model, runtime version and inputs: 26 cases for the classic tasks, 17 for EmbeddingGemma, 9 for Proofreader and 10 for Summarizer. Generated text must match byte for byte, and Google's releases generate different text, so each platform is compared with Google's wheel of its own version: the checked-in macOS references (Google's 1.1.0 wheel), references generated on the Linux and Windows runners (tool/test_modern_text.py) and for the gallery's emulator and simulator builds (tool/prepare_modern_text_reference.py), and Google's JavaScript on the same page for EmbeddingGemma in browsers. Fresh Flutter apps built from the packages as pub.dev ships them run the tasks in debug and release, alongside vision and audio. From the repository root:

make get models
make test_text

example/example.dart classifies a sentence from the command line: dart run example/example.dart "What a great film" in this package.

EmbeddingGemma, Proofreader and Summarizer are tagged modern-text: they need the generative models (make models downloads them), and dart test --tags modern-text runs them alone. The gallery's text pages demo them.