Your Medical Documents. Your Device. Your Privacy.
A privacy-first iOS application that brings the power of AI to your personal medical records—
100% offline, zero data leaves your device.
MediVault is a fully on-device Retrieval-Augmented Generation (RAG) system designed for personal medical document management. Scan your prescriptions, lab results, and medical records, then ask questions in natural language—all powered by AI running entirely on your iPhone.
"What was my blood pressure at my last checkup?"
"Show me my cholesterol levels over time."
"What medications was I prescribed in January?"
MediVault retrieves relevant information from your scanned documents and generates accurate, cited answers—without ever connecting to the internet.
| Feature | Description |
|---|---|
| 🔒 100% Offline | No network calls, no cloud storage, no telemetry. Your medical data never leaves your device. |
| 📷 Smart Document Scanning | Scan multi-page documents using your camera with automatic edge detection. |
| 🖼️ Photo Import | Import existing photos of medical documents from your library. |
| 🧠 On-Device AI | Powered by Qwen 2.5-1.5B running locally via llama.cpp—no API keys needed. |
| 🔍 Semantic Search | Find information using natural language, not keywords. Ask questions like you would ask a doctor. |
| ✅ Grounded Answers | Every response is validated against your actual documents with cited sources. |
| 🛡️ Medical Safety Guardrails | The AI is explicitly designed to never provide medical diagnoses or treatment advice. |
| 💬 Multi-Turn Conversations | Context-aware follow-up questions for deeper exploration of your records. |
MediVault implements a complete RAG pipeline optimized for mobile:
┌─────────────────────────────────────────────────────────────────────┐
│ MediVault Architecture │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ 📱 UI Layer (SwiftUI) │
│ ├── ChatView - Conversational interface │
│ ├── DocumentsTab - Document management │
│ └── ScannerView - VisionKit document capture │
│ │
│ 🧠 RAG Orchestrator │
│ ├── EmbeddingService - CoreML MiniLM (384-dim vectors) │
│ ├── VectorStore - SQLite + cosine similarity search │
│ ├── Phi4MiniService - Qwen 2.5-1.5B via llama.cpp │
│ ├── PromptBuilder - Context-aware prompt engineering │
│ ├── GroundingValidator- Claim verification against sources │
│ └── SafetyFilter - Medical safety guardrails │
│ │
│ 📄 Ingestion Pipeline │
│ ├── VisionOCRService - Apple Vision accurate text recognition │
│ └── TextChunker - Semantic chunking with overlap │
│ │
└─────────────────────────────────────────────────────────────────────┘
graph LR
A[User Question] --> B[Embed Query<br>MiniLM 384d]
B --> C[Vector Search<br>Top 3 Chunks]
C --> D[Build Prompt<br>Context + Query]
D --> E[LLM Generate<br>Qwen 2.5-1.5B]
E --> F[Validate<br>Grounding Check]
F --> G[Safety Filter]
G --> H[Display Answer<br>with Citations]
- Xcode 26.2+
- iOS 26.2+ device (iPhone only, simulator not recommended due to ML models)
- ~1GB storage for models
- Hugging Face CLI for model download
-
Clone the repository
git clone https://github.com/yourusername/MediVault.git cd MediVault -
Download the LLM model (~986MB)
mkdir -p MediVault/Resources/Models hf download \ enacimie/Qwen2.5-1.5B-Instruct-Q4_K_M-GGUF \ "qwen2.5-1.5b-instruct-q4_k_m-00001-of-00001.gguf" \ --local-dir MediVault/Resources/Models/💡 Tip: Install Hugging Face CLI with
pip install huggingface_hub -
Open in Xcode
open MediVault.xcodeproj
-
Build and run on your iOS device
On first launch, MediVault will:
- Load the CoreML embedding model
- Initialize the SQLite vector database
- Load the Qwen LLM into memory
This may take 15-30 seconds depending on your device.
- Navigate to the Documents tab
- Tap Scan Document to use your camera, or Import from Photos for existing images
- The app will:
- Extract text using OCR
- Split into semantic chunks
- Generate embeddings
- Store in the local vector database
- Navigate to the Chat tab
- Ask any question about your medical history
- View the AI's response with:
- Confidence indicator (green/orange/red)
- Source count showing how many documents were referenced
- Sources button to view the exact text snippets used
- "What was my hemoglobin level?"
- "When was my last vaccination?"
- "Summarize my visit to Dr. Smith"
- "What medications am I currently taking?"
| Component | Technology | Purpose |
|---|---|---|
| UI Framework | SwiftUI | Declarative, modern iOS UI |
| State Management | @Observable |
Swift 5.9+ observation framework |
| Concurrency | Swift Actors | Thread-safe services |
| Database | GRDB.swift | SQLite with Swift-native API |
| Embeddings | CoreML + MiniLM | 384-dimensional text vectors |
| Tokenization | swift-transformers | HuggingFace tokenizers |
| LLM Inference | swift-llama-cpp | Optimized on-device inference |
| OCR | Apple Vision | Accurate text recognition |
| Document Scanning | VisionKit | Native camera-based scanning |
MediVault is designed with privacy as the foundational principle:
| Aspect | Implementation |
|---|---|
| Data Storage | All data stored locally in app sandbox |
| Network Access | Zero network calls—completely offline |
| ML Inference | All models run on-device using CoreML and llama.cpp |
| Medical Safety | AI explicitly cannot provide diagnoses or treatment advice |
| No Analytics | No telemetry, no crash reporting, no user tracking |
⚠️ Disclaimer: MediVault is a document retrieval tool, not a medical device. Always consult healthcare professionals for medical decisions.
- Model: MiniLM-L6 (float16 CoreML)
- Dimensions: 384
- Sequence Length: 128 tokens
- Compute: CPU/GPU/Neural Engine
- Model: Qwen 2.5-1.5B Instruct (Q4_K_M quantization)
- Context: 4096 tokens
- Batch Size: 256
- Output: Structured JSON with citations
💡 Model Evolution: Initially developed with Phi-3 Mini 3.8B (4-bit), later migrated to Qwen 2.5-1.5B for improved performance on medical terminology and better structured JSON output compliance.
- Algorithm: Brute-force cosine similarity (optimized with Accelerate)
- Threshold: 0.5 minimum similarity
- Results: Top 3 most relevant chunks
- Chunk Size: 500 characters
- Overlap: 50 characters
- Boundary Detection: Sentence-aware splitting
MediVault/
├── App/
│ └── MediVaultApp.swift # App entry point
├── Features/
│ ├── Chat/ # Conversational UI
│ ├── Embedding/ # CoreML embedding service
│ ├── Generation/ # LLM inference
│ ├── Home/ # Tab navigation
│ ├── Ingestion/ # OCR & document import
│ ├── RAG/ # Core AI orchestration
│ │ ├── Grounding/ # Answer validation
│ │ ├── Orchestrator/ # Query coordination
│ │ ├── Prompt/ # Prompt engineering
│ │ ├── Protocols/ # Service interfaces
│ │ └── Safety/ # Medical guardrails
│ └── VectorDB/ # Vector storage
├── Resources/
│ └── Models/ # ML models (see setup)
└── Assets.xcassets/ # App assets
- iPad support with split-view
- Document categories and tagging
- Export reports as PDF
- Apple Watch companion for quick queries
- Multi-language OCR support
- iCloud backup (encrypted)
Contributions are welcome! Please feel free to submit a Pull Request.
- Fork the project
- Create your feature branch (
git checkout -b feature/AmazingFeature) - Commit your changes (
git commit -m 'Add some AmazingFeature') - Push to the branch (
git push origin feature/AmazingFeature) - Open a Pull Request
This project is licensed under the MIT License - see the LICENSE file for details.
- GRDB.swift - A toolkit for SQLite databases
- swift-transformers - Swift implementation of HuggingFace tokenizers
- swift-llama-cpp - Swift bindings for llama.cpp
- Qwen - The Qwen team for the open-source LLM
Built with ❤️ for privacy-conscious healthcare