A new demo on Hacker News showcases an AI system for home buying that takes a hardline stance against hallucination: it won't present any finding unless it can point to an exact quote from the source material. The project, called AskVetta, demonstrates this approach using a property at Meridian Avenue in San Jose as its test case.

Citation-First Architecture

The system appears to be built around a fundamental constraint: every piece of information it surfaces must be verifiable against input documents. Rather than relying on the AI to self-correct after generating confident nonsense, the architecture reportedly rejects conclusions that lack direct source support before they ever reach the user.

Why This Matters for Developers

For builders working with LLMs in high-stakes domains like real estate, legal, or healthcare, this represents a different architectural philosophy. Instead of layering evaluation and guardrails on top of generated outputs, AskVetta seems to constrain what the model can say based solely on what's explicitly present in the source material.

Trade-offs Worth Considering

The approach likely trades some coverage for reliability. A system that only speaks when it has direct quotes will inevitably miss inferences or connections that require interpretation beyond exact text matching. Whether that's an acceptable trade-off depends heavily on use case requirements.

Real Estate as a Proving Ground

Home buying is a smart domain to demonstrate this approach. Property decisions involve concrete factsβ€”square footage, zoning restrictions, comparable salesβ€”that have definitive answers in source documents. The San Jose demo suggests the team is targeting buyers who want auditable reasoning rather than confident recommendations.

Key Takeaways

  • Citation-gated architecture prevents hallucination at the inference level rather than correcting it afterward
  • Real estate's document-heavy nature makes it a natural fit for quote-verified AI outputs
  • The trade-off between coverage and reliability is an explicit design choice, not an afterthought

The Bottom Line

This isn't necessarily the right approach for every application, but it's a legitimate architectural pattern that deserves more attention. If you're building in high-stakes domains where errors have consequences, constraining outputs to verifiable claims might be worth trading some of that "AI magic" for genuine trustworthiness.