Last updated

Adopt your existing notes and make them AI-queryable

You already have notes. Maybe years of them, in another app or a folder on disk. This guide takes that pile and turns it into a connected, searchable, AI-answerable workspace: the same files you came in with, now linked together, visible as a map, and grounded enough that an agent can answer real questions from them.

Importing is where this starts, but it is not where the value is. A folder of loose Markdown is just a folder. The payoff comes from reconnecting those notes, seeing their shape, and letting AI read them. That is the path below.

This assumes you have already brought files in, or are about to. If you have not, start with Importing and adopting and come back here.

Step 1: Get the files in

Point Undra at your Markdown and it shows up as notes right away. Nothing is converted, your filenames stay as they are, and the files remain plain Markdown in a folder you own. The exact mechanics of staged drops, where notes land, and what Undra adds when you save are covered in Importing and adopting.

Why this matters here: everything after this step builds on the imported notes being real, editable items in your workspace. Get them in first, then enrich.

Not ready to copy files in?

If you would rather leave the folder where it sits on disk and work on it in place, a folder portal does that without importing. The rest of this guide still applies to the notes inside it.

Notes imported from another tool usually arrive disconnected. Each one knows nothing about the others, even when they are clearly related. That is the orphan problem, and it is the single biggest gap between a folder of files and a workspace.

The fix is wikilinks. Type [[ in a note and link it to another item by name, and you get a two-way connection: the note you link to gains a backlink back, and the graph gains an edge. You do not have to design folders or tags first. You link as you read.

Why do this before anything else AI-related: wikilinks are the structure the rest of Undra runs on. Backlinks come from them, the map in the next step is drawn from them, and an agent following a note’s links picks up the related material around it. The syntax (aliases, embeds, where links work) lives in Wikilinks and backlinks.

Link as you reread, do not schedule a cleanup

You will not reconnect a whole archive in one sitting, and you do not need to. Drop a [[wikilink]] whenever you notice two notes belong together, and let the connections accumulate as you actually use the notes.

Step 3: See the shape with the graph

Once links exist, open the graph. Every note becomes a node and every wikilink an edge, so your whole import shows up as a map instead of a list.

This is where you spot two things a file tree hides: clusters of notes that clearly belong to one topic, and orphans drifting unconnected at the edges. The clusters tell you where your real subjects are. The orphans tell you what still needs a link. Coloring patterns you care about (with groups) and saving a view (as a preset) is covered on the graph page.

Why this is a step and not a detour: the graph turns reconnection into a visible, finishable job. You link, you look, you see the gaps shrink. It is the feedback loop that makes Step 2 worth doing.

Step 4: Build the semantic index once

By default, AI search over your workspace matches keywords. Semantic search lets it match by meaning, so a question about “money coming in” can find a note titled “Revenue” even with no shared words. That power comes from an index you build once, by hand. It does not build itself the first time, and search will not answer by meaning until it exists.

This step stays on your machine

Building the index runs entirely on your own computer: a small local model turns your notes and plans into vectors, which are stored locally in .undra/embeddings.sqlite. No API key, no provider, nothing uploaded. Turning the feature on and the one-time build are covered in Setting up AI.

Why bother: once the index exists, an agent searching your imported library finds the right note by what you mean, not just the words you happened to type, which is most of what makes the next step feel like it actually read your work.

Step 5: Ask an agent a question grounded in the import

Now point an agent at the material you just brought in. Ask it something that needs your real notes to answer, not something it could guess from general knowledge: “what did I conclude about X across these notes?” or “summarize everything here that touches Y.”

Because the agent reads your actual files (the note in front of it, the items you select, the things they link to) the answer is grounded in your import instead of the open web. The wikilinks from Step 2 widen what it can pull in, and the index from Step 4 helps it find the right notes by meaning. How to set an agent’s permissions and keep it read-only while you trust it is covered in Agents.

Start read-only

For a first pass over freshly imported notes, an agent that can read your workspace but cannot change anything is the safe default. Let it answer and summarize before you ever let it propose edits. The permission switches are on the Agents page.

That is the whole arc: files in, links made, shape seen, index built, questions answered. The same Markdown you started with, now something you can interrogate.

Where to go next