Add TFLite lightweight local AI solution with user-imported models

[Copilot]

Implements 方案 D: TFLite-based on-device AI with user-imported models for embedding, ASR, and OCR. No pre-bundled models—users download/import as needed, keeping the app lightweight (50-100MB total depending on choices).

Core Services

• LocalEmbeddingService: TFLite inference for paraphrase-multilingual-MiniLM (25MB, 384d, 50+ languages). O(1) vocab lookups via Map.
• LocalSpeechRecognitionService: Interface for user-imported Whisper ONNX (tiny 39MB / base 74MB)
• LocalOCRService: Supports Tesseract traineddata or PaddleOCR Lite (~8MB)
• LocalVectorStorageService: SQLite BLOB storage with brute-force cosine similarity (suitable for <10k notes)
• LocalAIModelManager: File picker-based model import with validation

Models

• LocalAIModelConfig: Model paths, status, metadata
• Extended LocalAISettings with modelConfig property
• LocalAIModelRecommendations: Download URLs for recommended models

UI

Model management section in LocalAISettingsPage:

• Per-model status (Ready/Not installed)
• Import/delete actions
• Storage statistics
• Recommendation dialogs with copy-to-clipboard download links

Usage

// Initialize embedding service with user-provided model
final embeddingService = LocalEmbeddingService();
await embeddingService.initialize(
  '/path/to/model.tflite',
  '/path/to/vocab.txt',
);

// Generate embeddings
final embedding = await embeddingService.generateEmbedding('Hello world');

// Store and search
await vectorStorage.storeEmbedding(noteId, embedding!);
final results = await vectorStorage.searchSimilar(queryEmbedding, topK: 10);

Dependencies

• tflite_flutter: ^0.10.4 (Apache-2.0)
• record: ^5.1.0

Limitations

• No LLM support (纠错、来源识别 not available)
• Brute-force search—consider dedicated vector DB for >10k notes

Original prompt

方案 D：TFLite 轻量级（全用户导入）

技术栈

• 嵌入: tflite_flutter (Apache-2.0) - paraphrase-multilingual-MiniLM (TFLite 25MB)
• ASR: sherpa_onnx 推理用户导入的 Whisper ONNX 模型
• OCR: 用户导入 Tesseract traineddata 或 PaddleOCR Lite 模型
• 向量存储: SQLite 暴力搜索

核心实现

1. 多语言嵌入 (TFLite)

模型: paraphrase-multilingual-MiniLM-L12-v2 TFLite 版

• 量化后 25MB, 384维
• 支持50+语言
• 用户在应用内下载或手动导入 .tflite 文件

class LocalEmbeddingService {
  Future<void> initialize(String userTflitePath, String vocabPath);
  Future<List<double>> generateEmbedding(String text);
}

2. ASR (用户导入模型)

用户导入 Whisper ONNX 模型 (tiny 39MB / base 74MB)：

class LocalSpeechRecognitionService {
  Future<void> initialize(String userWhisperPath);
  Future<String> transcribeFile(String audioPath);
}

3. OCR (用户导入模型)

选项 A: Tesseract traineddata

await FlutterTesseractOcr.extractText(
  imagePath,
  language: 'chi_sim+eng',
  args: {'tessdata': userTessdataPath},
);

选项 B: PaddleOCR Lite (~8MB, 中文识别更优)

// 使用 paddle_ocr_flutter 或 tflite_flutter 加载用户导入的模型

4. SQLite 向量存储

暴力搜索（适合 < 10k 笔记）：

// 数据库存储 BLOB 格式向量
CREATE TABLE note_embeddings (
  note_id TEXT PRIMARY KEY,
  embedding BLOB NOT NULL
)

// 搜索时遍历计算余弦相似度

依赖配置

dependencies:
  tflite_flutter: ^0.10.4
  sherpa_onnx: ^1.11.3
  flutter_tesseract_ocr: ^0.4.23  # 或
  paddle_ocr_flutter: ^0.1.0       # PaddleOCR
  record: ^5.1.0
  camera: ^0.11.0

模型管理 UI

在设置页提供：

• 嵌入模型: 推荐下载链接 + 手动导入
• ASR 模型: 推荐 Whisper tiny/base 下载 + 手动导入
• OCR 模型: Tesseract traineddata 或 PaddleOCR 模型导入

模型总大小: 50-100MB（取决于用户选择）

优势

• 最轻量: 无预装模型，用户按需下载
• 灵活性: ASR/OCR 模型用户自选
• 低内存: TFLite 量化模型占用小
• 商业安全: 全部开源许可

功能限制

• ❌ 无 LLM（不支持 AI 纠错、来源识别）
• ❌ 暴力搜索不适合大规模数据（>10k 笔记考虑方案 C）
• ✅ 支持语义搜索、相关笔记、简单分类

适用场景

• 低端设备
• 存储空间有限
• 不需要复杂 LLM 功能
• 快速原型验证

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[vercel]

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

Project	Deployment	Review	Updated (UTC)
thought-echo	Ready	Preview, Comment	Jan 2, 2026 0:49am

[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.}