Show HN: WebPizza – AI/RAG pipeline running in the browser with WebGPU

🚀 Live Demo: https://webpizza-ai-poc.vercel.app/

⚠️ Experimental POC: This is a proof-of-concept for testing purposes only. It may contain bugs and errors. Loosely inspired by DataPizza AI.

100% Client-Side AI Document Chat - No servers, no APIs, complete privacy.

Chat with your PDF documents using AI that runs entirely in your browser via WebGPU.

• 🔒 100% Private: All processing happens in your browser - your documents never leave your device
• ⚡ Dual Engine: Choose between standard WebLLM or optimized WeInfer (~3.76x faster)
• 🤖 Multiple Models: Phi-3, Llama 3, Mistral 7B, Qwen, Gemma
• 📄 PDF Support: Upload and chat with your PDF documents
• 🎯 RAG Pipeline: Advanced retrieval-augmented generation with vector search
• 💾 Local Storage: Documents cached in IndexedDB for instant access
• 🚀 WebGPU Accelerated: Leverage your GPU for fast inference

• Frontend: Angular 20
• LLM Engines:
  • WebLLM v0.2.79 (Standard)
  • WeInfer v0.2.43 (Optimized with buffer reuse + async pipeline)
• Embeddings: Transformers.js (all-MiniLM-L6-v2)
• PDF Parsing: PDF.js v5.4.296
• Vector Store: IndexedDB with cosine similarity
• Compute: WebGPU / WebAssembly

• Modern browser with WebGPU support (Chrome 113+, Edge 113+)
• 4GB+ RAM available
• Modern GPU (Intel HD 5500+, NVIDIA GTX 650+, AMD HD 7750+, Apple M1+)

1. Open chrome://flags or edge://flags
2. Search for "WebGPU"
3. Enable "Unsafe WebGPU"
4. Restart browser

Check your browser: https://webgpureport.org/

1. Select Engine: Choose between WebLLM (standard) or WeInfer (optimized)
2. Choose Model: Select an LLM based on your hardware capabilities
3. Upload PDF: Drop your document (first load downloads model ~1-4GB)
4. Ask Questions: Chat with your document using natural language

• ❌ No data collection
• ❌ No server uploads
• ❌ No tracking cookies
• ❌ No analytics
• ✅ 100% client-side processing
• ✅ Your data never leaves your device

See our Privacy Policy and Cookie Policy for details.

• Phi-3 Mini: ~3-6 tokens/sec
• Llama 3.2 1B: ~8-12 tokens/sec
• Mistral 7B: ~2-4 tokens/sec

• ~3.76x faster across all models
• Buffer reuse optimization
• Asynchronous pipeline processing
• GPU sampling optimization

The project includes vercel.json with optimal configuration for WebGPU and routing.

Ensure your hosting supports:

• SPA routing (all routes → index.html)
• Cross-Origin headers for WebGPU:
  • Cross-Origin-Embedder-Policy: require-corp
  • Cross-Origin-Opener-Policy: same-origin

1. Check browser version (Chrome/Edge 113+)
2. Enable chrome://flags#enable-unsafe-webgpu
3. Update graphics drivers
4. Test at https://webgpureport.org/

• Try a smaller model (Llama 1B, Qwen)
• Use WeInfer engine for 3.76x speedup
• Close other tabs/applications
• Check GPU isn't throttling

• Use smaller models
• Close other browser tabs
• Increase browser memory limit
• Clear browser cache and restart

This is a proof-of-concept project. Contributions, issues, and feature requests are welcome!

MIT License - See LICENSE file for details

Emanuele Strazzullo

• Website: emanuelestrazzullo.dev
• LinkedIn: linkedin.com/in/emanuelestrazzullo

• MLC LLM - WebLLM inference engine
• WeInfer - Optimized WebLLM fork
• Transformers.js - Browser ML library
• PDF.js - PDF parsing
• Hugging Face - Model hosting

Made with ❤️ by Emanuele Strazzullo