Mobile Search Client

A high-performance, local-first mobile application for document indexing and semantic search. This project allows users to search through personal documents (PDF, TXT, MD) using a combination of traditional keyword matching (BM25) and modern vector embeddings—all processed entirely on-device for maximum privacy.

Features

• Hybrid Search Engine: Combines Full-Text Search (FTS5) with Vector Similarity (RRF) using op-sqlite.
• Native PDF Extraction: High-performance text and page parsing via PDFBox (Android) and PDFKit (iOS).
• On-Device Embeddings: Uses onnxruntime-react-native to generate embeddings locally.
• Privacy First: No data leaves the device; all indexing and searching happen offline.
• Smart Chunking: Page-aware document splitting for precise search result navigation.

Prerequisites

• Node.js: 18.x or later
• React Native CLI
• Android: Android Studio & SDK (API 21+)
• iOS: macOS with Xcode (15+) and CocoaPods

Installation

1. Clone the repository:

   git clone <repository-url>
   cd client

2. Install dependencies:

   npm install
   # or
   yarn install

3. Install iOS Pods:

   cd ios && pod install && cd ..

Model Setup (Required)

The app uses the multilingual-e5-small model for generating text embeddings. To keep the repository lightweight, the model file is not included and must be downloaded manually.

1. Create the models directory if it doesn't exist:

   mkdir -p assets/models

2. Download the quantized ONNX model. You can use the following command (or download it manually from a source like Hugging Face):

   # Example: Downloading from a compatible source
   curl -L "https://huggingface.co/intfloat/multilingual-e5-small/resolve/main/model.onnx" -o assets/models/model_quantized.onnx

   Note: Ensure the filename is exactly model_quantized.onnx.

3. Verify you have the following files in assets/models/:

   • model_quantized.onnx (The file you just downloaded)
   • tokenizer.json (Already in the repo)
   • tokenizer_config.json (Already in the repo)

Running the App

Android

npx react-native run-android

iOS

npx react-native run-ios

Search Modes

• BM25: Classic keyword-based search. Best for finding exact terms and names.
• Semantic: Vector-based search. Best for finding concepts and related meanings even without exact word matches.
• Hybrid: Combined ranking (Reciprocal Rank Fusion). Provides the most balanced results by merging both strategies.

Project Structure

• android/ & ios/: Native modules for PDF extraction and platform configurations.
• src/services/: Core logic for Database, Embedding, and Search services.
• src/screens/: UI components for searching, document management, and results.
• assets/models/: Tokenizer and ONNX model files.

License

MIT