VPSask & VPask

The problem

Enterprise agreements and policy manuals run to hundreds of pages. The answer to a practical question — am I paid overtime on a Sunday, what leave can I take when my partner has a baby — is usually spread across a clause, a schedule and a policy. General-purpose chatbots answer confidently and cite nothing, and a wrong answer about pay or leave has real consequences.

What we built

Two independent, non-profit question-and-answer sites. VPSask covers the Victorian Public Service Enterprise Agreement 2024 and related policies; VPask covers Victoria Police policy and guidance. People ask in their own words, and every statement in the answer links to the page it came from. The built-in document viewer opens at that page with the quote highlighted. When the sources don't support an answer, the assistant says so and suggests better search terms instead of guessing.

Both sites state plainly that they are not official Victorian Government or Victoria Police services and are not legal advice. Every answer sends people back to the cited source.

Engineering notes

• One multi-tenant Next.js codebase and one Cloud Run service. The request host selects the knowledge base, branding and copy.
• A Python retrieval pipeline: query expansion and model-assisted rewrites, BM25 retrieval, rank fusion, scope filtering and a local reranker.
• Grounding is enforced on the server. Each claim must be anchored to a verbatim quote, and quantity and negation checks run on top. A verifier model removes unsupported claims, and the visible answer is rebuilt from what survives. The model never writes citations or page numbers; the server builds the page links.
• An evaluation suite measures retrieval, answer grounding, answer safety and security headers. Each release is deployed without traffic and promoted only after a smoke test passes.
• Cost guardrails: a daily question budget, rate limiting, bounded concurrency and scale-to-zero hosting.