The AI core: how HOWI decides what's true

A confident, well-written wrong answer is worse than no answer — especially about a Wi-Fi password, a checkout time, or how to work the thermostat. This shipped the engine and the rule that make HOWI refuse to guess about the things that matter.

The problem

A general-purpose AI model will fill any gap with a plausible-sounding answer, including for facts specific to one property that it has no way to actually know.

What shipped

• A multi-pass pipeline for host AI — intent, plan, answer, and an optional review-and-retry pass — with an internal eval harness to catch regressions before they reach anyone
• Async compression of runtime facts, so long conversations about devices, stays, and policies stay fast without losing operational detail
• Retrieval infrastructure (pgvector-based) so AI can pull relevant context instead of relying only on the current message
• A verified-context rule for guest AI: every answer is classified as verified (in the actual configuration), an unverified suggestion (clearly labeled as such), a general local recommendation, or unknown — with "ask the host" as an honest fallback, never a guess

Why it matters

This is the rule underneath every guest answer HOWI has given since: specific facts about a property are only ever stated as true when they're actually configured. Everything since — voice, personality, local recommendations — has had to work inside this constraint, not around it.

Availability

Shipped December 22–26, 2025, and enforced on every guest conversation since. No benchmark or accuracy percentage is published — the constraint is architectural (a rule enforced in the prompt and pipeline), not a measured model behavior.