Building Lumi: what it knows, what it won't guess
Lumi only knows what you've told it. That sounds like a limit. It's actually the most important design decision in the product.
The hardest design decision in tentus wasn't what Lumi should do. It was what Lumi should refuse to do.
Most assistants are trained to be helpful at any cost. Ask them something they don't know, they'll fabricate. That works for a chat toy. It's a disaster for a thinking tool, where you'll be making decisions on top of what Lumi tells you.
The rule
Lumi only knows what you've told it.
If a fact lives in your workspace — a note, a meeting transcript, a connected source — Lumi can surface it, cite it, and reason about it. If it doesn't, Lumi says so. No paraphrased general knowledge, no plausible-sounding fill-ins, no confident summaries of things it can't see.
What that buys you
Trust. The first time Lumi tells you it doesn't know something, you breathe out. Because now you know that when it does answer, the answer is grounded in your own captures.
It also buys you correctness in a sneaky way. A lot of the wrong answers from AI tools come from filling gaps. Remove the incentive to fill gaps, and the wrong-answer rate plummets.
How we built it
Retrieval-augmented generation is the load-bearing piece. Every Lumi response is forced through a retrieval step against your workspace, with the retrieved sources passed to the model alongside the strict instruction: if the sources don't support the answer, say so.
Underneath that, the embedding pipeline runs continuously. Every captured source gets chunked, embedded, and indexed within seconds. The latency floor on briefings is set by how fast that pipeline runs, not by how big your workspace gets.
What it won't be
Lumi won't pretend to be your friend. It won't crack jokes. It won't volunteer opinions. It's a thinking layer, not a personality. The work is the point.