Anonymous, so no memory
No accounts means no saved history. Tenants keep a copy by printing or copying.
England only
Housing law differs in Wales, Scotland and Northern Ireland. A signpost beats wrong guidance.
The phone’s voice, not a server’s
Every phone already has dictation and read-aloud. Server speech would cost about 13 times the answer itself. No audio reaches the service, so there is no new privacy risk.
The Talk to us experiment is the exception. Its audio goes to Google, and is not stored.
Questions first, at a cost
Asking first should fit the advice to the case. But 21 of 82 live test chats never reached advice.
Vector search, then hybrid again
Version 1 logs showed the 0.5 similarity cut-off never filtered anything. Vector search alone also missed exact terms, such as Section 21 versus Section 8. Version 2 fuses full-text and vector search.
One slow query, fixed
A retrieval query fell from 2,911 ms to 136 ms. That is 21 times faster.
Measured with EXPLAIN ANALYZE, May 2026.
Tenants only, not landlords
Landlords already have access to legal help. Tenants often do not. Landlord questions get a pointer elsewhere. Some questions sit in between, and those can go wrong.
Accuracy before speed
For legal guidance, a wrong answer is worse than a slow one. Accuracy is checked against an evaluation set; it is not guaranteed.
One function, one database
One Supabase Edge Function and Postgres with pgvector, not a team of agents and a separate vector store. Fewer parts means fewer things to break.
