mediapipe_text 0.1.3
mediapipe_text: ^0.1.3 copied to clipboard
Google's MediaPipe text tasks for Flutter: text classification and embeddings, language detection, EmbeddingGemma, proofreading and summarization.
0.1.3 #
- The README has an installation section using
flutter pub add. - With
mediapipe_core0.1.3, downloads of models and native libraries wait out short network failures.
0.1.2 #
Release 0.1.2.
0.1.1 #
Adds Google's NOTICE at the package root.
0.1.0 #
The first release of this rewrite, replacing Google's
mediapipe_text 0.0.1 (May 2024) with a new API;
MIGRATION.md maps it.
- Text Classifier, Text Embedder and Language Detector on Android, iOS, macOS, Linux, Windows and the web, through Google's official runtimes.
- On 32-bit Android (armeabi-v7a) the Text Classifier, Text Embedder and Language Detector run on the CPU; EmbeddingGemma, the Proofreader and the Summarizer need 64-bit Android, since their models take more memory than a 32-bit process holds.
- EmbeddingGemma, as a
TextEmbedderwithTextModels.embeddingGemma, on the same six platforms, with Google's prompt formatting on every runtime:embed(text, {formatContext})withEmbeddingType,TextRoleandTextFormatContext. - Proofreader and Summarizer on Android, iOS, macOS, Linux and Windows, with
streamed updates (
proofreadStream,summarizeStream). Browsers have neither task (upstream-issues.md UP-034), which the capability queries report. - Every task is one class with the same API on every platform, created with
await Xxx.create(options)and released with an idempotentdispose().mediapipe_text.dartis the only app library. - Options are flat on core's
TaskOptions(model,modelPath,modelBytes,delegate) with Google's settings. Results are plain values on core'sClassifications,EmbeddingandMediaPipeCategory, andTextEmbedder.cosineSimilarityis static. - Per-task capability queries (
queryTextClassifierCapabilities()and the others) report the supported delegates and why any other is unavailable. Invalid classifier settings fail in Dart on every platform, andcreaterefuses the GPU, which no text task runs on. TextModels.byNamenames every pinned model forhooks.user_defines.mediapipe_text.models, whichdart run mediapipe_core:bundle_modelsbundles into the app.model:uses the bundled copy and downloads at run time only whenModelStore.allowDownloadsis true; app-supplied paths and bytes work too.