02. Architecture

02. Architecture & Tech Stack

DIRD+ is a Tauri v2 desktop app: a Rust native shell hosting a WebView that runs the React frontend. All logic — UI, AI inference, database, encryption, LLM — runs in-process on the device. There is no server and no network dependency after the initial model download.

┌──────────────────────────── Tauri process (Rust shell) ────────────────────────────┐
│                                                                                     │
│   WebView (frontend)                          Rust backend (src-tauri/)             │
│   ┌─────────────────────────┐                 ┌─────────────────────────────────┐   │
│   │ React 18 + TypeScript    │   Tauri IPC    │ crypto.rs   Argon2id KDF        │   │
│   │ Zustand stores           │ ◄───commands──►│ db.rs       SQLCipher (AES-256) │   │
│   │ Konva canvas             │                │ llm.rs      llama.cpp inference  │   │
│   │ ONNX Runtime Web (WASM)  │                │ models.rs   custom-model registry│   │
│   │ jsPDF reports            │                └─────────────────────────────────┘   │
│   └─────────────────────────┘                                                       │
│                                                                                     │
│   AI inference (ONNX) runs in the WebView via WebAssembly.                           │
│   The database, encryption and LLM live in the Rust backend.                        │
└─────────────────────────────────────────────────────────────────────────────────────┘

Frontend (src/)

Tech	Version	Role
React	18.3	UI framework
TypeScript	5.7 (strict)	Static typing
Vite	6	Build + HMR
Tailwind CSS	3.4	Styling
Radix UI	—	Accessible primitives (Dialog, Tabs, Select, Switch, Slider)
React Router	6	SPA routing
Zustand	5	Global state (config, canvas, patient)
Konva / react-konva	10	Multi-layer annotation canvas
jsPDF + jspdf-autotable	2.5 / 3.8	Client-side PDF reports
i18next	24.2	i18n (Spanish / English)
Vitest + happy-dom	—	Test suite

AI & inference

Tech	Role
ONNX Runtime Web 1.23	WebAssembly inference engine (detection + segmentation), runs in the WebView
OpenCV.js	Optic-disc refinement
llama.cpp (llama-cpp-2 crate)	Optional in-process LLM for report-prose polishing — no remote inference

Rust backend (src-tauri/)

Module	Responsibility
crypto.rs	Argon2id key derivation (OWASP 2025 params: m=64 MiB, t=3, p=4)
db.rs	SQLCipher (AES-256) database via rusqlite (bundled-sqlcipher-vendored-openssl); raw key passed via PRAGMA key = x'…'
llm.rs	Local LLM: curated model catalog, on-demand download with progress events, per-family chat templating, in-process generation
models.rs	Custom-model registry under <user_data>/models/ (install / list / uninstall / set-active)

Storage

• Live database: SQLite + SQLCipher (encrypted at rest). Access layer in src/lib/db-sql/ (TableShim, mappers, migrator). Schema in src-tauri/migrations/0001_init.sql.
• Legacy: src/lib/db/ contains the old Dexie/IndexedDB schema, kept only for one-time migration of pre-SQLite installs. New installs never use it.
• Export: .dird containers (ZIP, AES-256-GCM). See 06. Imports & Exports.

Project layout

src/
├── components/   canvas/ patients/ upload/ reports/ settings/ onboarding/ demo/ ui/
├── lib/          ai/ analysis/ clinical-guidelines/ db-sql/ db/ export/ pdf/ classes/ api/
├── stores/       config-store, canvas-store, patient-store
├── i18n/         es / en locales
└── App.tsx       router + parallel init
src-tauri/        crypto.rs, db.rs, llm.rs, models.rs, migrations/