Experimental browser-native local RAG using Chrome's built-in Gemini Nano runtime.
ChromeRAG explores a simple question:
If Chrome is becoming a local AI runtime, what new local-first software architectures become possible?
This repository contains two experimental builds:
app/- ChromeRAG Connectors, a localhost web app for local files, folders, URLs, workspaces, evidence retrieval, and optional SharePoint / Microsoft Graph connector scaffolding.extension/- ChromeRAG Sidecar, a Chrome extension side panel that can index pages as you browse and query indexed evidence locally.
The project started as a practical experiment after reports that Chrome had downloaded a local LLM onto users' machines. Rather than treat that as a curiosity, this repo asks what happens if we use the browser itself as an AI runtime.
The core idea:
browser content + local storage + retrieval + Chrome Prompt API = local-first intelligence layer
This is not intended to be production-ready. It is an architectural exploration of browser-native AI, private RAG, and local enterprise knowledge augmentation.
- Local folder and file indexing
- URL ingestion
- Workspace-based evidence stores
- BM25-style retrieval
- Chrome
LanguageModel/ Prompt API synthesis - Prompt context budgeting for Gemini Nano
- Citation-style evidence snippets
- Markdown and styled HTML exports
- Microsoft Graph / SharePoint connector direction
- Chrome extension side panel prototype
- Recent Chrome / Chrome Canary with built-in AI / Prompt API enabled
- Node.js 20+
- Windows/macOS/Linux desktop
The local LLM API must be exposed in the browser:
'LanguageModel' in window
await LanguageModel.availability()cd app
npm install
$env:PORT=8795
npm startOpen:
http://localhost:8795
chrome://extensions
→ Developer mode
→ Load unpacked
→ select extension/
Experimental. Expect Chrome Prompt API behaviour, model availability, context limits, and feature flags to change.
MIT