ChatVault 🔒

Your WhatsApp history, indexed locally with AI.

ChatVault is a local-first semantic search engine for WhatsApp exports. It uses Rust and WebAssembly to run a BERT neural network directly in your browser.

No servers. No data egress. 100% Private.

demo_video.mp4

🚀 The Problem

WhatsApp's native search is strictly keyword-based. If you search for "recommendations for sushi", it won't find the message where your friend said "we should go to that japanese place on 5th".

ChatVault solves this by generating vector embeddings for your chats locally. It understands meaning, not just keywords.

💻 Try it out

Option 1: Live Demo (Recommended)

The fastest way to test the engineering is via the Vercel deployment. The Wasm module loads directly in your browser.

👉 Launch ChatVault

Option 2: Running Locally

If you want to inspect the Rust systems layer or modify the code:

1. Clone the repo

   git clone https://github.com/marcoshernanz/chat-vault.git
   cd chat-vault

2. Install dependencies

   # Ensure you have Rust installed (rustup)
   cargo install wasm-pack
   cd web && npm install

3. Run Development Server This command compiles the Rust core to Wasm and starts the Next.js server concurrently.

   npm run dev:all

🛠️ Tech Stack (The "How")

This project pushes the browser to its limits, combining high-performance systems programming with modern React patterns.

Core (Systems Layer)

• Rust: Memory-safe business logic and vector storage.
• Candle (Hugging Face): Runs a quantized BERT model (all-MiniLM-L6-v2) inside Wasm for generating embeddings.
• WebAssembly (Wasm): Compiles Rust to binary for near-native performance in the browser.
• Hybrid Search Algorithm: Implements a custom scoring system combining Cosine Similarity (Vector) + Keyword Matching (BM25-style) for maximum accuracy.

Frontend (Application Layer)

• Next.js 16 (App Router): Modern React framework.
• Web Workers: Offloads heavy AI inference and indexing to a background thread to keep the UI at 60fps.
• IndexedDB: Caches the model weights and vector index for instant subsequent loads.
• Tailwind CSS v4: Fluid, responsive UI.

⚡ Performance & Engineering

1. Hybrid Search Implementation

Pure vector search can sometimes miss exact names or specific dates. ChatVault implements a weighted hybrid approach in Rust:

// Simplified logic from core/src/lib.rs
// We combine semantic meaning with exact keyword matches
let hybrid_score = (vector_score * 0.5) + (keyword_score * 0.5);

// Heuristics for noise reduction:
if content_len < 30 && keyword_score < 0.01 {
    // Penalize short messages that don't match keywords 
    // (Reduces noise like "ok", "cool", "lol")
    hybrid_score *= 0.4; 
}

2. Zero-Blocking Architecture

Running a Neural Network in JS usually freezes the browser. ChatVault uses a Worker-first architecture:

1. Main Thread: Handles Drag & Drop and UI rendering.
2. Worker Thread: Loads the 90MB Model, tokenizes text, runs inference, and performs the vector search.
3. Communication: Uses generic message passing for non-blocking UI updates.

3. Smart Parsing

Includes custom Regex parsers for both iOS and Android WhatsApp export formats, handling multi-line messages and system notifications automatically.

🔒 Privacy Note

This application is 100% offline-capable. When you drop your WhatsApp text file, it is processed entirely within your device's RAM/WebAssembly memory. No data is ever sent to a server. You can disconnect your internet after the model loads and it will still work.

👤 Author

Marcos Hernanz

• GitHub
• LinkedIn
• X (Twitter)

---

Built with ❤️ in Madrid. Looking for Summer 2026 roles in SF.