Product · Knowledge Graph
A graph of your ideas, built while you write
Korely reads every note as you save it and pulls out the names, projects, and ideas that come up across your vault. Notes that share them become related on their own, with no tagging discipline. The same graph powers GraphRAG, so when you or your AI ask a question, the answer comes from the whole network of your notes.
Entities
Every note read, every name found
A small named entity recognition model runs on your CPU and scans every note as you save it. Think of it as a librarian skimming each page as it lands on the desk, marking who and what showed up. The text never leaves your machine.
The model picks up five kinds of things:
- People (like "John Smith" or "Sarah")
- Organisations (a customer name, OpenAI, a vendor)
- Places (Berlin, Potsdam, an office, a city)
- Projects (the Q3 roadmap, a feature codename, a campaign)
- Recurring concepts (GraphRAG, MCP, local-first, pricing)
Variants collapse to the same node. "John" and "John Smith" point at one entry. The extraction takes seconds per note and runs in the background while you keep writing. More on the entity model →
From the founder
On three calls in two weeks the same customer talked about change request 22, change request 25, and change request 25 again. By the third call I had mixed them up: I was answering about CR-22 while the client was asking about CR-25. The kind of mistake that becomes very hard to make once the graph holds every mention of every name on its own. The graph never gets confused: each request is its own node with its own history. After that week I stopped trusting my memory for cross-call details and started trusting the vault.
Related notes
Notes link themselves while you write
Open any note and the sidebar shows the related ones: every note in your vault that mentions at least one of the same entities as the one in front of you. Ranked by overlap, deduplicated, refreshed each time you save.
Picture a detective's pinboard. Every time two notes mention the same name, a red string appears between them on its own. You did not have to remember the connection or type it. The note about last week's call links to the proposal draft and the contract review because they all mention the same customer.
You can still type wikilinks like [[note name]]
if you want explicit edges. They become hard links in the graph.
The auto-link layer keeps working on top.
Graph view
Navigate the vault as a map
The /notes view in Korely is the
graph itself. Nodes are notes, edges are shared entities, node size
grows with how many connections a note has. The clusters that appear
on the canvas are the topics you actually write about, drawn from
your real notes rather than a template.
Hover a node to see the title. Click to open the note. Filter by an entity (a customer name, a project, a concept) and the canvas collapses to that subset. It is how you go from "I remember writing about this in two places" to actually finding both places in a second.
GraphRAG
Answers that walk the graph, not just the words
Plain retrieval hands the AI a stack of paragraphs that contain the keywords from your question. That is good for "what is the definition of X" and weak the moment the question spans several notes: "summarise everything we decided about EU pricing last quarter."
GraphRAG (graph augmented retrieval) hands the AI both: the paragraphs that match the words and the notes connected to the topic through the entity graph. "EU pricing" pulls in every note that mentions Europe or the pricing rework, even when the exact phrase is missing. The answer comes back with the full picture. Hallucinations go down because the model has more of the actual source to work from.
Korely's built-in chat (Pro) uses this pipeline. So does the MCP
tool korely_get_related, which is
what Claude, Cursor, or ChatGPT call when you wire Korely as an MCP
source. The graph is the layer they all read.
More on MCP →
Scale
Big vault, small wait
Entity extraction runs in a background scheduler that polls every five seconds for new or modified notes. On an idle vault it sleeps. On an active one it throttles so typing stays smooth. New notes show up in the graph within seconds of being saved.
The graph lives in SQLite, in the same database file that holds your notes. Queries like "find related to this note" or "list every note that mentions Sarah" return in milliseconds on a vault of ten thousand notes. The whole thing is a single file you can copy, back up, or move between machines.
Frequently asked
Do I have to type wikilinks to get the graph? +
No. The graph is built on its own. Korely reads every note as you
save it and links notes that mention the same things. Wikilinks
like [[note name]] still work if
you type them, but you do not have to.
How does Korely find the entities? +
A small named entity recognition model runs on your CPU and scans every note. It picks up people, organisations, places, projects, and recurring concepts. The model never sends your text anywhere. See the entity extraction page for the details.
What is GraphRAG? +
GraphRAG is retrieval that uses the entity graph alongside vector search. When you ask Korely or your AI a question, the answer pulls from the notes that match the keywords and from the notes connected to the topic through shared entities. Better recall on multi-step questions, fewer made-up answers.
Can I see the graph visually? +
Yes. The /notes view is the graph itself. Nodes are notes, edges are shared entities, node size grows with the number of connections. Hover to see the title, click to open the note, filter by entity to collapse the view onto a single topic.
Does the graph slow Korely down on large vaults? +
Entity extraction runs in the background and throttles itself when the vault is idle. New notes are processed in seconds. The graph lives in the same SQLite file as your notes and queries return in milliseconds on vaults with ten thousand notes or more.
The graph builds itself. You keep writing.
Free forever for the local vault. Pro for cloud sync and AI when you want them.