b6fe000 docs(build-log): notes for 2026-10-09 and 2026-10-101d428ea docs(tasks): Rey search + context review04894f5 fix(rey): short follow-ups stay on the page; voice the first ~200 chars, 30 s relaya5a2e34 fix(rey): retell Del's first person as Del, stream-safeac58580 fix(rey): answer from the passages specifically, retell Del's first person as Del48d90c2 chore: drop committed __pycache__, ignore it5d7165e feat(rey): meaning search, a full-site index, and page awareness56c1737 docs(lessons): scroll triggers and page-scoped tours6f19314 feat(companion): Rey is scoped to every page, and narrates in reading order536c45e docs(tasks): Rey companion reviewb45a5a9 fix(companion): live voice signs what he actually says, and only the openinge3af956 feat(companion): Rey, a free-roaming guide on every page, with his own voiceRey got a body and a voice. He's a small drone who flies around every public page: he greets
you, tours the page you're on section by section, points at what he's talking about, and reacts
when you poke, pet or fling him. He speaks with an open-source voice (Kokoro, "Rey · warm")
that runs on the homelab. Every scripted line is pre-recorded, so a page view costs nothing,
and only a chat answer is voiced live. In the 3D console he flies beside your ship and narrates
each world.
I caught two bugs on the homepage. The pricing line spoke before the Who line, because a
section taller than two phone screens can never be "half visible". And every page ran the same
tour. Lines now fire at mid-screen, in reading order, and each page tours its own sections. The
18 case studies weren't loading Rey at all; now they are.
Then the part you can't see: how Rey finds his facts. The index went from 66 passages to 388,
covering every case study section by section, plus the services, guides and about pages. Search
now matches meaning, not just words: a small embedding model on the homelab reads each question
in about 15 ms. "My product spreadsheet is a mess" now lands on the catalog work. Rey also knows
which section you're looking at, so "what was the result there?" means that project.
What it still can't do: the 3B model sometimes summarises a passage instead of quoting its
numbers. Search is no longer the weak link; model size is.