mp_text

Text task constructors, results, streaming APIs, and backend availability.

API#

Task	Methods	Result types
`LanguageDetector`	`detect(text)`	`LanguageDetectorResult`
`TextClassifier`	`classify(text)`	`ClassificationResult`
`TextEmbedder`	`embed(text, formatContext: ...)`	`EmbeddingResult`
`TextProofreader`	`proofread`, `proofreadStreaming`	`TextProofreaderResult`, `TextProofreaderChunk`
`TextSummarizer`	`summarize`, `summarizeStreaming`	`TextSummarizerResult`, `TextSummarizerChunk`

Create each task with TaskName.create(TaskNameOptions(...)) and release it with close(). All options contain BaseOptions; classifier and embedder options also contain their shared mp_core option groups.

LanguageDetector#

Returns ordered BCP-47 language predictions and probabilities. A real-browser integration test runs the official language detector model and verifies the model digest path.

TextClassifier#

Returns classification heads with stable category indices, scores, labels, and display names. Shared ClassifierOptions provide allowlists, denylists, maximum results, thresholds, and locale selection.

TextEmbedder#

Returns floating-point or quantized embedding heads. Enable L2 normalization when cosine similarity is the intended comparison and keep the choice consistent between indexed and query embeddings.

TextProofreader#

proofread returns corrected text plus an ordered list of unchanged, inserted, and deleted segments. proofreadStreaming emits text chunks and includes the final correction list on the completed chunk.

final proofreader = await TextProofreader.create(
  TextProofreaderOptions(
    baseOptions: BaseOptions(
      modelAsset: ModelAsset.path('models/proofreader.litertlm'),
    ),
    maxTokens: 2048,
  ),
);

TextSummarizer#

TextSummarizerMode.tldr requests a short paragraph. TextSummarizerMode.keyPoints requests a key-point list and is the default. Both blocking and streaming methods are exposed.

Backend availability#

Task	Web	C runtime	Android	iOS
Language detection	Implemented	Implemented	C packaging in validation	Planned
Text classification	Implemented	Implemented	C packaging in validation	Planned
Text embedding	Implemented	Implemented	C packaging in validation	Planned
Text proofreading	No upstream API	No upstream API	Native constructor tested	Native constructor tested
Text summarization	No upstream API	No upstream API	Native constructor tested	Native constructor tested

Unsupported combinations fail during task creation with MpStatus.unimplemented.

Task instances own native or JavaScript resources. Reuse them across calls and close them at the feature lifecycle boundary.

Only one proofreader or summarizer operation may be active on a task instance. Cancelling a Dart subscription stops delivery to that subscriber; it does not cancel generation in the platform SDK. Wait for the native operation to finish before starting another call or closing the task.