When you ask Ret a question, the answer is only as good as the notes it finds first. An AI given the wrong three notes will write a confident, well-formatted answer about the wrong thing. So most of the work behind Ret happens before any AI is involved: finding the right notes, quickly, on your own computer.
Here's how that works, step by step, with as little jargon as we can manage.
Step 1: notes become meaning
When you save something, ReticulOS quietly prepares it for searching.
First it splits the note into passages, following its structure rather than chopping every so many words. A table stays in one piece. So does a list, or a callout. That way a passage always makes sense on its own, which matters later when it's shown to an AI.
Then each passage is turned into a long list of numbers that captures what it means. This is called an embedding. Two passages about the same idea end up with similar numbers, even if they don't share a single word. “Dr Lee, Thursday at 4” and “the dentist appointment” land close together.
The model that does this is called EmbeddingGemma, and it runs on your computer. The first time ReticulOS needs it, it downloads the model, about 200 MB. After that, preparing notes and searching them needs no connection at all.
Pictures, PDF pages and recordings get the same treatment, from the pictures and sounds themselves. That's why a photo of a receipt with no title can still match a search for “receipt”, and why a recording can be searched by describing what was said. Reading PDF pages and recordings this way is part of Pro.
Step 2: two searches, combined
When you ask a question, it's turned into numbers the same way, and ReticulOS runs two different searches at once.
A search by meaning compares your question's numbers with every passage's numbers and finds the closest. This is what finds “the dentist” when your note says “Dr Lee, Thursday at 4”.
A search by keywords looks for the actual words. It's what makes sure an exact phrase, a product code or a name is never missed, which is exactly where searching by meaning alone tends to stumble.
The two lists of results are merged into one, with a small extra nudge for notes you saved recently, since recent notes are more often the ones you mean.
All of this runs on your computer. On our test laptop, turning a question into numbers takes about 25 milliseconds.
Step 3: an answer with sources
Only now does the AI come in. Ret takes the best-matching notes, gives each one a number, and asks the AI to answer your question from those notes and to cite them by number.
What you see is an answer with small numbered citations through it, and a list of the notes it used. Click any citation and Reticulum opens on that exact note, highlighted, even if it's in a different collection or a day in your day book.
Why it works this way
We could have sent your whole library to an AI service and let it sort things out. It would have been simpler to build. It would also have been slower, more expensive per question, and a lot less private.
Doing the searching on your computer means it works offline, costs nothing per question, and keeps everything you haven't asked about where it is. Asking the AI to cite numbered notes means you never have to take an answer on trust: every claim points back to something you saved, and you can check it in a click.
Getting the best answers
A few habits make a real difference to what Ret finds:
- Give important notes a title. The title is searched along with the note, so a good one helps the right note surface, and makes its citation easy to recognise.
- Keep related things in one collection, and ask that collection. A focused set of notes gives a focused answer.
- Use All collections when you're not sure where something is. It searches everything, including the day book.
- Ask in your own words. Because of the search by meaning, you don't need to remember how you phrased the note.
