feat: implement ONNX Runtime + ObjectBox local AI stack (方案 C)

[Copilot]

Implements on-device AI capabilities using ONNX Runtime for inference and ObjectBox for vector search, enabling offline semantic search, embedding generation, OCR, and ASR.

Dependencies

• onnxruntime: ^1.16.0 - ONNX inference (MIT)
• objectbox: ^4.0.0 + objectbox_flutter_libs + objectbox_generator - Vector DB (Apache-2.0)
• sherpa_onnx: ^1.11.3 - ASR (Apache-2.0)
• google_mlkit_text_recognition: ^0.13.0 - OCR (MIT)
• record: ^5.1.0, camera: ^0.11.0 - Input capture

New Services (lib/services/local_ai/)

• ONNXInferenceService - Model loading, caching, inference execution
• TokenizerService - WordPiece tokenization with CJK support
• EmbeddingService - Text→vector using paraphrase-multilingual-MiniLM-L12-v2 (384d)
• VectorSearchService - HNSW-indexed search via ObjectBox, similarity caching
• ASRService - Whisper/Paraformer speech recognition (placeholder, needs model files)
• OCRService - ML Kit text recognition with multi-script support

ObjectBox Entities (lib/models/objectbox/)

• NoteVector - Note embeddings with HNSW index
• SearchHistory - Query tracking
• SimilarNotesCache - Precomputed similarity relationships

Usage

// Generate embedding
final embeddingService = EmbeddingService();
await embeddingService.initialize();
final result = await embeddingService.embed("Hello world");

// Vector search
final vectorSearch = VectorSearchService();
await vectorSearch.initialize();
final results = await vectorSearch.search("find notes about AI", topK: 10);

Setup Required

1. dart run build_runner build - Generate ObjectBox code
2. Download model files to assets/models/

Original prompt

方案 C：ONNX Runtime + ObjectBox

技术栈

• 嵌入: onnxruntime (MIT) - paraphrase-multilingual-MiniLM-L12-v2 (多语言)
• ASR: sherpa_onnx (Apache-2.0) - Whisper/Paraformer
• OCR: google_mlkit_text_recognition (MIT)
• 向量存储: ObjectBox Vector Search (Apache-2.0)

核心实现

1. 多语言嵌入模型

推荐: paraphrase-multilingual-MiniLM-L12-v2

• 支持50+语言（含中英文）
• 量化后 118MB, 384维
• 下载：HuggingFace/Xenova

备选: text2vec-base-multilingual (中文优化, 110MB, 768维)

2. ONNX 推理引擎

class ONNXInferenceService {
  Future<void> loadModel(String modelId, String path);
  Future<Map<String, OrtValue>> runInference(String modelId, inputs);
}

3. Tokenizer

class TokenizerService {
  Future<void> loadVocab(String vocabPath);
  TokenizedInput encode(String text);
}

4. ObjectBox 向量搜索

@Entity()
class NoteVector {
  @HnswIndex(dimensions: 384)
  List<double> embedding;
}

// 搜索
vectorBox.query()
  .nearestNeighborsF32('embedding', queryVector, topK)
  .find();

5. ASR/OCR

• ASR: sherpa_onnx (Whisper tiny 39MB 或 Paraformer)
• OCR: ML Kit 自动下载中文模型

依赖配置

dependencies:
  onnxruntime: ^1.16.0
  objectbox: ^4.0.0
  sherpa_onnx: ^1.11.3
  google_mlkit_text_recognition: ^0.13.0
  record: ^5.1.0
  camera: ^0.11.0

生成 ObjectBox 代码:

dart run build_runner build

优势

• 模型灵活可替换（HuggingFace 生态丰富）
• ObjectBox 向量搜索性能优于暴力搜索（适合大规模笔记）
• 全部开源许可，商业安全
• 多语言嵌入模型统一处理中英文

注意事项

• 需要自行实现 Tokenizer（提供 WordPiece 示例代码）
• 无 LLM，不支持 AI 纠错和来源识别
• 适合对模型选择有特定需求的团队

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Review	Updated (UTC)
thought-echo	Ready	Preview, Comment	Jan 3, 2026 1:39am

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}