OPAX: Open Parliamentary Accountability eXchange

Ask what Australian politicians actually said, see how they voted, and follow who funds them.

opax.com.au

518,685 speeches from five parliaments | 13,867 recorded divisions | 199,233 donations classified by industry | 6 standing investigations

What OPAX does

OPAX puts the parliamentary record in front of the reader and lets them question it. Every answer is written only from retrieved passages and cites them; if the record is thin, the answer says so.

• Ask the record. Grounded, cited answers over Hansard from the Commonwealth, New South Wales, Victoria, South Australia and Queensland, plus Senate committee hearings, bills and official government statements. Filter by speaker, party, state, topic or years. Answers stream in as they are written.
• Search. Hybrid keyword and semantic search with the same filters, a per-search answer rail, and exports (CSV, BibTeX, RIS).
• Encyclopedia. An entry for every parliamentarian, party, donor and topic: quick facts, portrait, voting record, declared interests, expenses, and the money map isolated on that entity.
• Money map. A 3D map of the largest donors, ringed by industry around the parties they gave to, with the words layer showing who speaks on each industry's debates. Federal, Queensland and Victorian disclosures.
• Standing investigations. Six reports (climate and energy, gambling, housing, immigration, First Nations, media ownership) regenerated from the knowledge box, pairing the money with the words.
• Explore. Time machine, the record quiz, the ledger, who owns which debate, words per dollar, then versus now.

Architecture

Browser (static SPA, hash router)
   │
   ▼
Cloudflare Worker  portal/src/index.ts     /api/search  /api/ask (JSON or SSE)  /api/resource  /api/topic ...
   │
   ▼
Progress Agentic RAG knowledge box         speeches, bills, official statements, division records; topic labels and summaries from enrichment tasks

There is no live database behind the site. Structured data (donations, votes, expenses, interests, grants, contracts) stays relational in a SQLite corpus on the data box and reaches the site as static JSON exported by scripts/export_*.py.

MIGRATION-ARAG.md is the runbook: knowledge box identity, model pins, corpus inclusion rules, cost gates, platform quirks learned the hard way.

Repository layout

Running the site locally

cd portal
npm install
printf 'ARAG_KB_TOKEN=...\n' > .dev.vars   # the knowledge box token; zone and KB id are in wrangler.jsonc
npx wrangler dev                              # http://localhost:8787

Run the portal tests with npm test from portal/. Some suites read Python fixtures that need openpyxl: in a fresh worktree run uv sync --frozen at the repo root first (or set OPAX_PYTHON), otherwise those tests skip, and with CI set they fail.

Deploy with npm run deploy from portal/ — not npx wrangler deploy. The
npm script runs scripts/stamp_assets.mjs first, which content-hash stamps the
?v= on /app.js and /style.css in index.html; those two are served
immutable for a year, so an unstamped deploy strands returning visitors on the
old bundle. See docs/HARDENING.md.

The committed public/money-map.js must equal a fresh build:

npx esbuild graph/index.ts --bundle --minify --format=esm --target=es2022 --outfile=public/money-map.js
node graph/smoke-test.mjs

UI reference

Component language, control recipes, homepage structure and remaining consolidation work. The homepage uses /; Ask and Search share /ask with view=search selecting Search. Workbench and prototype review assets are local-only and excluded from deployment.

Data pipeline

The corpus lives in ~/.cache/autoresearch/parli.db on the data box. Python tooling is managed with uv:

uv sync
uv run python -m parli.ingest.arag_sync --tables speeches --full   # push to the knowledge box
uv run python scripts/generate_reports.py                                          # regenerate the six reports
uv run python scripts/ask_harness.py                                               # 24 grounded questions against production

Reports retrieve original records across the corpus, with date windows rather than a Hansard-only topic filter. Each report includes a dedicated non-speech evidence question. Generated summaries and bill registry cards are excluded; original bill text, explanatory memoranda and other primary records remain eligible. Speech counts and charts use the current topic-labelled speech catalog; the financial charts retain the separately audited scripts/report_stats.json snapshot.

Set REPORT_STRUCTURED_WRITER=codex to use the installed Codex CLI for structured figures, party positions and introductions. Narratives still use the knowledge box’s configured writer. Run uv run python scripts/validate_reports.py before publishing to check citations against the live corpus.

Environment variables are documented in .env.example. Never commit .env or portal/.dev.vars.

Data sources

Licensing and coverage per source are recorded in the docs/ notes and the *_DATA_SOURCES.md surveys.

Contributing

Work on a branch and open a pull request; main is deployed manually and always matches production. Useful places to start:

1. Data coverage: ingestion for WA, Tasmania, the NT and the ACT parliaments.
2. The record: harness questions in scripts/ask_questions.json, and prompt or retrieval improvements in portal/src/index.ts.
3. The site: the static front end is plain HTML, CSS and JavaScript with no build step.

Origins

OPAX grew out of an autoresearch experiment, pivoted to parliamentary transparency. The first version ran as a FastAPI and Next.js stack on a single server; it was retired in September 2026 in favour of the Worker and knowledge box described above.

License

OPAX is licensed under the GNU Affero General Public License v3.0. You are free to use, modify and distribute it, but modifications to the web service must also be made available as open source.

Built with data from OpenAustralia, TheyVoteForYou, the AEC, the state parliaments and electoral commissions, and the Guardian and ABC.