DavidVeritasr/recall

Veritasr Recall

Offline, browser-native document intelligence. No cloud. No egress. Runs on anything.

What it is

A fully client-side RAG system. Documents are indexed offline, embeddings run in-browser via WebGPU or WASM fallback, and a small ONNX language model answers questions locally. Nothing leaves the device.

Built for environments where cloud access is unavailable, restricted, or unacceptable.

How it works

1. 1.A Python pipeline (separate repo) ingests documents and produces a flat JSON index with BGE embeddings
2. 2.The browser app loads that index and handles all retrieval and inference locally
3. 3.No backend. No API calls. No data leaving the tab.

Requirements

• —Chrome or Edge recommended for WebGPU acceleration
• —Firefox works via WASM CPU fallback - significantly slower
• —No installation. No account. No internet after initial load.

Tech

• —Transformers.js for in-browser query embedding
• —BGE embeddings, flat JSON index format
• —ONNX language models via WebGPU / WASM fallback
• —Vite + React

License

Apache 2.0

Built by Veritasr LLC - Guilford, CT