Why Nightdesk doesn't use RAG (and why that's fine for a guestbook)
A few people asked why Nightdesk doesn't do retrieval like most AI products. Here's the honest tradeoff.
A few people have asked why Nightdesk doesn't use retrieval — the standard "chunk your documents and search them" approach most AI products reach for by default. Fair question, so here's the honest answer: I looked at it and decided it was solving a problem I don't have.
A guestbook is not a knowledge base
Retrieval earns its complexity when you're searching across thousands of documents, or a corpus too big to fit in a model's context window. A single listing's guestbook is a few thousand words. Every modern model can just... read the whole thing, every time, no search step, no risk of retrieving the wrong chunk and confidently answering from the wrong context.
So that's what Nightdesk does. The whole guestbook goes into the prompt on every message. Simpler to reason about, nothing to misretrieve, and it means the answer is always grounded in the complete guide, not a fragment of it.
The tradeoff, honestly
It costs a bit more per message than a lighter retrieval setup would, since the full guidebook gets resent each time rather than cached once and searched. For a single listing at realistic guest-message volumes, that cost is still small — cents per month, not dollars, at the free-tier size. It only starts to matter at a scale most individual hosts will never hit.
I'd rather ship the simpler, more reliable version now and revisit if real usage ever demands otherwise, than add retrieval complexity pre-emptively for a problem that doesn't exist yet.
Why this matters to you as a host
Practically: it means when a guest asks something oddly specific — a weird combination of two things you mentioned in different parts of the guestbook — the answer can actually connect them, because the model saw the whole document, not just the one paragraph a search step guessed was relevant.
If you want to see it answer against your actual guestbook instead of a demo one, the free tier takes about five minutes to set up.