example/README.md

flutter_gemma_embeddings example #

flutter_gemma_embeddings is the runtime-agnostic on-device text-embedding pipeline for flutter_gemma — tokenization, the background-isolate worker, and pooling/normalization, over an EmbeddingForwardPass seam. It does not ship a concrete backend itself (since 2.0.0); pair it with an engine package that provides one, e.g. flutter_gemma_litertlm's LiteRtEmbeddingBackend (Gecko / EmbeddingGemma .tflite via the LiteRT C API — dart:ffi on the 5 native platforms, LiteRT.js on web). Register the backend once at startup, then embed text and feed the vectors into any RAG vector store.

import 'package:flutter/widgets.dart';
import 'package:flutter_gemma/flutter_gemma.dart';
import 'package:flutter_gemma_embeddings/flutter_gemma_embeddings.dart';
import 'package:flutter_gemma_litertlm/flutter_gemma_litertlm.dart';

Future<void> main() async {
  WidgetsFlutterBinding.ensureInitialized();

  // The backend comes from an engine package; the tokenizers from this one.
  await FlutterGemma.initialize(
    embeddingBackends: [LiteRtEmbeddingBackend()],   // flutter_gemma_litertlm
    embeddingTokenizers: [GemmaEmbeddingTokenizers()],
  );

  // Install an embedding model (downloads + sets it active). The model and its
  // tokenizer are separate downloads.
  await FlutterGemma.installEmbedder()
      .modelFromNetwork('https://example.com/embeddinggemma.tflite', token: 'hf_...')
      .tokenizerFromNetwork('https://example.com/sentencepiece.model', token: 'hf_...')
      .install();

  // Create the embedding model and embed text.
  final embedder = await FlutterGemma.getActiveEmbedder();
  final vector = await embedder.generateEmbedding('Gemma runs on-device.');
  print('embedding dim: ${vector.length}');

  await embedder.close();
}

Pair this with a RAG vector store (flutter_gemma_rag_qdrant on native, flutter_gemma_rag_sqlite for web) to build on-device retrieval. A full runnable app lives in the flutter_gemma example.