Architecture

Overview

Browser Extension (Chrome)
├── content.js          — article scanning, sidebar UI
├── background.js       — service worker, pipeline orchestration
├── popup.js            — scan trigger, popup UI
└── src/
    ├── llm.js          — dual-model AI extraction
    ├── api.js          — Congress.gov + GovTrack lookups
    ├── votesmart.js    — VoteSmart client (via proxy)
    ├── lookup.js       — politician name resolution
    ├── keywords.js     — topic keyword extraction
    └── topic-match.js  — bill relevance matching

Backend Proxy (Node.js / Render)
└── server/
    ├── index.js                  — Express, CORS, rate limiting
    └── providers/
        ├── claude.js             — Anthropic Claude extraction
        ├── mistral.js            — Mistral extraction
        ├── _shared.js            — shared prompt + parser (single source of truth)
        ├── congress.js           — Congress.gov proxy
        └── votesmart.js          — VoteSmart JWT auth + proxy

Pipeline

When you click Scan This Page:

1. content.js extracts article text and politician names via regex
2. popup.js sends an analyze message to the service worker
3. background.js calls extractArticleAnalysis() in src/llm.js
4. src/llm.js sends { articleText } to two proxy endpoints in parallel:
   • POST api.liarsledger.com/api/claude/extract
   • POST api.liarsledger.com/api/mistral/extract
5. Both models return { summary, main_topics, figures[] } — claims are compared via Jaccard similarity
6. background.js resolves politician names against a 3,606-key dictionary
7. src/api.js fetches sponsored/cosponsored legislation and GovTrack roll-call votes via api.liarsledger.com/api/congress/*
8. src/votesmart.js fetches interest group ratings and vote history via api.liarsledger.com/api/votesmart/*
9. Results are stored in browser.storage.session and displayed in the sidebar

Key design decisions

Why a backend proxy? All API keys (Claude, Mistral, Congress.gov, VoteSmart) live server-side. The extension ships with zero secrets. This also enables rate limiting and future freemium tier enforcement.

Why two AI models? Two independent models from different companies with different training data reduces systematic bias. Agreement = higher confidence. Disagreement = both interpretations surfaced.

Why GovTrack instead of Congress.gov for votes? The Congress.gov senate-vote and house-vote endpoints return 404 for the 119th Congress — they are in beta and not yet available. GovTrack provides complete roll-call vote history at no cost.

Why session storage for caching? browser.storage.session clears when the browser closes, preventing stale legislative data from persisting. Cache hits are logged in the debug panel.