006: Reverse RAG: cross-page context via encrypted localStorage

Reporter: fxhp Date: 2026-03-03 Priority: medium Status: open Affects: Browser extension, embedded widget

Description

When a user browses multiple pages on the same site, the machine learning model should be able to reference content from previously visited pages. Currently each page is isolated: the model only knows the current page's content.

Goal

Build a reverse RAG system where page summaries are stored in encrypted localStorage (per-domain), and automatically included in the system prompt for new pages. This gives the model a growing knowledge base about the site as the user browses.

Architecture

After generating the intro message for a page, extract a brief summary (title, key topics, 2-3 sentences)
Store summaries per-domain in vault-encrypted localStorage: uncloseai-pageknowledge-{domain}
On new pages, include the last N page summaries in the system prompt as context
Summaries should be compact (under 200 tokens each) to avoid context overflow
Use the existing fitPageContent() budget system to allocate space for cross-page context

Privacy

All cross-page data must be encrypted via the vault, same as chat history. The user controls their data. No data leaves the device except through explicit LLM requests.

Dependencies

Site journey tracking (implemented in storage.js)
Vault encryption (already working)
fitPageContent() context budgeting (already working)

Notes

This is the foundation for making uncloseai "unstoppable" on any site. The model gets smarter the more pages you visit, building a personal knowledge graph of the site, all stored locally and encrypted.

1.7 KiB Raw Blame History