Search it, then ask it
Three ways to find something, in increasing order of how vague you are allowed to be. All of them run on your own hardware by default.
Keyword search
⌘K (or Ctrl+K) opens search from anywhere. It is a Lucene index over every
note, returning ranked hits with the matching phrase highlighted in context.
The overlay searches more than notes: categories, collections and even settings pages are matched, so “retention” takes you to the Trash setting rather than making you remember which tab it lives on.
If the index is rebuilding or the server is unreachable, search does not simply stop working — it falls back to matching over the notes already cached in your browser and labels the results as partial. Cruder, but never a dead end.
Semantic search
Lucene finds the words you typed. Semantic search finds the note you meant.
Papyra chunks each note, embeds the chunks locally, and stores the vectors alongside the index. Searching for “that thing about pricing per user” will surface a note that says “we settled on per-seat” without either phrase overlapping. It is the right tool when you remember the shape of an idea but not its vocabulary.
Ask your notes
The spark icon opens an assistant that answers from your vault. It retrieves the relevant chunks, hands them to a model, and streams the answer back — with the notes it used cited underneath, as links you can open and check.
It is a conversation, not a one-shot query: earlier turns are kept and sent along, so “what about the second one?” resolves. Conversations are saved per account, renameable, and deletable; a transcript of your notes never leaves the account that owns it.
When it cannot answer, it says why. “No model installed” and “your API key was rejected” call for very different actions from you, and Papyra distinguishes them instead of returning an unexplained blank.
The model runs on your machine
The default is Ollama, alongside Papyra, on your hardware. Nothing is sent anywhere. Papyra offers three one-click installs in plain language rather than making you choose model names:
| Tier | Model | Download | RAM needed |
|---|---|---|---|
| Small | llama3.2:1b | 1.3 GB | 2 GB |
| Balanced | llama3.1:8b | 4.7 GB | 8 GB |
| Best | mistral-nemo:12b | 7.1 GB | 12 GB |
A model is several gigabytes, so the install reports real byte progress rather than an indeterminate spinner — an unexplained ten-minute wait reads as a hang.
If you would rather use a hosted model, an admin can point the instance at OpenAI or Anthropic from Settings. Chat and embeddings are configured independently, so answering with a hosted model while keeping search embeddings entirely local is a supported combination. API keys are write-only over the API: the server reports whether a key is stored, never what it is.