- You paste your notes folder into an AI sidebar, ask for a plan, and the plan shows up in forty seconds. Now ask the question nobody asks while it works: where did those notes go?
- This article follows the real route of a question, a document and an image through this kind of tool, using the vendor's public documentation. You will see why a local search does not make the answer local.
- You leave with a ten-minute inventory: what leaves your machine, where it goes, and the rule to apply before you paste a note.
You paste your notes folder into the AI sidebar, ask for a plan, and the plan arrives in forty seconds. It is impressive. Then a question crosses your mind: where are those notes now?
An AI sidebar is a relay. It receives your request, reshapes it, sends it through a model, then gives you the answer back. Several steps in between can touch your documents, and none of them shows on screen. The tool is not necessarily doing something wrong. The problem is that almost nobody reads the list of what leaves the machine.
[ANECDOTE DIANE: the moment you realised a confidential document had travelled to a tool without you knowing, and what you changed afterwards.]
A question goes to the model you picked
The first route is the simplest, and usually the only one people imagine. Your question goes to the model you selected: ChatGPT, Claude, Gemini or another. For Triplo AI, whose version 7.0 documentation is dated 20 October 2026, the privacy page is explicit: the complete prompt goes to the endpoint you chose. The complete prompt is not just your sentence. It is the tool’s instruction, the passages from your documents retrieved to answer, the ongoing conversation and your text.
That is the part that matters. The model provider matters less than the content. Everything the tool assembles around your question travels with it, and that content is often far wider than what you typed.
A local model changes this. With Ollama or LM Studio running on your machine and hosted models switched off, the question does not leave your PC. But that is a choice you have to make and verify, not a default. One hosted model selected by mistake, and the route changes silently.
Retrieved passages leave with the question
This point changes the picture most for notes. A sidebar that “searches your documents” usually does two things. It finds the relevant passages on your disk, then pastes them into the request sent to the model. The search can be entirely local while the answer is not.
Triplo describes this plainly. Its search index runs on your machine, but the passages it selects are attached to the prompt, and they go to the provider as soon as you query a hosted model. A tool can say your notes stay with you, then send several paragraphs of your file with every question if you ask it for a summary.
The rule I have followed since: what is searched locally is not necessarily processed locally. I check the two columns separately.
Media and links follow their own route
An image, a recording or a URL does not follow the same path as a text. In the same documentation, images and audio go to the vendor’s vision and transcription endpoints, even when the chosen model is local. There is no local fallback in this version. A YouTube link goes to a transcript service, and a web page is fetched with an ordinary request from your computer.
This gives a simple distinction. A screenshot of your bank statement has no business in a sidebar, even with a local model, until you know where it goes. For each kind of content, the question is not “is this tool safe?” It is “what kind of content am I putting in, and by which route?”
What stays on your disk
The route does not end at sending. There is also what remains on your PC, and that is where the technical details matter.
Saved conversations stay in a local file, which is good news. Model API keys, however, live in a configuration file. The same documentation says none of these credentials are encrypted at rest. The licence key also appears in backups and in the debug data the tool copies.
In practice, if you back up your Documents folder or send a bug report to support, you can carry credentials without noticing. That is not a scandal. It is one line to add to your backup routine: check what a tool writes in plain text before you share a folder.
A simple habit completes that check: before you paste a screenshot, crop it to what the question needs. A full screen shows your tabs, your messages and your desktop, far more than the question requires. The same goes for debug reports: before you send one to a support team, open it and read it. If a line contains a key, remove it before you send anything.
What the marketing pages say, and what the list says
Presentation pages talk about trust, control and “your data belongs to you”. They describe intent, almost never route. I have started reading them as a promise, then looking for the table.
The table exists at Triplo. For each action, it gives the destination and what is sent. It also spells out two things presentation pages often skip. Licence validation sends the key, a device fingerprint, the processor model, the username and the platform. Device sharing, when enabled, sends prompts, instructions and automations to a service run by the vendor.
None of this makes the tool dangerous. It makes it readable. A tool that states in black and white what it transmits is easier to use knowingly than one that promises trust without detail.
A ten-minute inventory
Reading every contract does not help much. What works is listing what you actually do with your notes, then checking the destination of each action in the tool’s own documentation.
- Open the tool’s privacy or security page. Look for a table or list titled something like “what leaves your machine”. If there is none, that is already information.
- Write down the five actions you repeat most with your notes: pasting text, attaching a document, asking for a summary of a folder, attaching an image, pasting a link.
- For each one, note the destination and what is sent. An empty box is a question to ask the vendor, not a reassuring answer.
- Sort your content into three levels: never in the tool, in the tool with a hosted model, in the tool with a local model only.
- Test your chosen setting with a note that does not matter. Check that the answer comes from the model you selected, and that the setting survives a restart.
Ten minutes is enough for one sidebar. It is not enough for a whole system, but it gives you a rule, and a rule is faster to apply than the doubt that returns each time you paste a text.
Here is what one filled line looks like, for summarizing a folder with Triplo. The destination is the endpoint you chose. The content sent is the complete prompt, including the passages retrieved from your documents. The level is hosted model only if the folder can already be shared, local model otherwise. Your own table will differ, and that is fine: it depends on what you actually write, not on what the tool promises.
The three levels also guard against the most common slip: a setting that changes with the menus. The model selected at the top of the screen can replace the one you meant to use, and the route follows the model, not your intention. The Triplo privacy page suggests a fix: pin a local model to every sensitive prompt, so the choice no longer depends on what is shown at the top of the screen.
What this changes for your notes
You do not have to stop using AI on your notes. You have to know which notes you hand over. A plan built from public course notes is no problem. A client file, a password you wrote down by mistake in a note, or your personal reflections on your work deserve different handling, often a local model, or no tool at all.
The right order stays the one from the Create lesson: you pick the output before the tool, and AI stays an execution layer, not the source of truth. Here, add one question before the first request: where does this text go?
If you want to keep context across your notes, files and actions, the logic is the same. Context usually breaks when information moves somewhere without anyone noting it, and an AI sidebar is one more place it can move.
Start today with one line: the name of the tool you use most with your notes, followed by the first action you repeat every week. Then build the table for that action only. One checked action is worth more than five tools you skimmed. Repeat the check after each major update, because a new version can change the route.
Sources
Figures and features checked on 5 November 2026.
See the 4 sources for this article
- Triplo AI, documentation, 20 October 2026, What’s New in 7.0
- Triplo AI, documentation, 20 October 2026, Privacy & Security
- Ollama, official site
- LM Studio, official site