How FairWork Mate AI stays current with FWC and Federal Court decisions

Every general-purpose AI model has a training-data cutoff. ChatGPT's most current public model was trained on data up to a fixed date months ago. After that date, every new Fair Work Commission decision, every section amendment, every Annual Wage Review — invisible. It will still answer your question; it just won't know what it doesn't know.

That's fine for “explain how casual loading works generally.” It's dangerous for “is the way my employer is doing this lawful as of right now?”

Every weeknight we run a scheduled job that:

1. Fetches the FWC's decisions RSS, the Federal Court's daily judgment list, and the Fair Work Ombudsman litigation page.
2. Extracts new decisions, sends each through a summarisation prompt that produces structured fields (facts, outcome, employer implication, employee implication, tags).
3. Stores the structured record in our database with a slug for the public summary page.
4. Generates a vector embedding for the summary and stores it in pgvector.

By the time you ask FairWork Mate AI a question the next morning, last night's decisions are already in the retrieval index.

When you ask a question that's likely to benefit from case-law (we triage every query), we:

1. Embed your question with the same model used on the case summaries.
2. Cosine-similarity search against all 260+ cases.
3. Return the top three above a relevance threshold.
4. Inject those three cases into the AI's system prompt with explicit instructions to cite at least one in the answer.

The AI doesn't hallucinate cases — it can only reference what we've actually loaded into its context. If retrieval finds nothing relevant, the AI says so honestly rather than making something up.

Two reasons. First, search engines rank for general traffic — they don't know which decisions actually apply to a worker's specific circumstances. Vector similarity does. Second, the FWC and Federal Court don't expose a decent search API; getting structured access to recent decisions is the moat.

We also publish each case as a public summary at /cases — both because users like reading the source, and because it gives Google a reason to index our coverage.

State workplace tribunals (Industrial Relations Commission of NSW, etc.) aren't in the corpus. We'll add them as we expand. Pre-1 January 2024 decisions are also limited — the corpus prioritises recency over completeness because the rate of change in workplace law is accelerating (right to disconnect, NDAs in workplace sex discrimination, non-compete reform, Payday Super, PPL extension).

If you want a deeper-cut historical search, the AustLII archive is free and excellent.

Ask /advisor something specific to your situation. Notice the citations in the answer. Click through to the case summary. Then ask ChatGPT the same question and notice what's missing — the link to the actual decision.