Your data and intelligence

Build intelligence around what you know.

Bring your documents, project notes, reviewed examples, and public research into a local Velora project. Open Intelligence in the desktop app to manage sources, review useful knowledge, and prepare training data for models you control.

1. Choose a project and import your sources

1. Add a local project from Projects, then open Intelligence and select it.
2. Open Sources to upload a text or Markdown file, paste source text, or collect a public page URL. Text files can be up to 500 KB; CSV, JSON and JSONL files are imported as source text.
3. Inspect the saved source text, revision, and hash. Give the source a useful title and reference so you can trace it later.

For a private document, export its text or Markdown and import that file. You can also use the Velora Sources extension to review visible browser text and save it to the selected local project. Google Docs sign-in and PDF extraction are separate from this capture flow.

Sync a server into Sources

1. In Intelligence → Sources, choose Add a server and enter its address — an internal wiki’s API, a documentation server or a JSON endpoint, on the internet or on your company network.
2. Add a token if the server needs one. It is kept in your operating system’s keychain; the project database stores only a reference to it, and Velora refuses to follow a redirect while a token is attached.
3. Choose Connect and sync. JSON, Markdown, text, CSV, XML and HTML are converted to source text. Each later sync adds a revision only when the content changed.

Capture browser text directly into Sources

Download Velora Sources for Chrome or Edge. It requires Velora 3.0.0-beta.13 or newer.

1. Extract the ZIP, then use Load unpacked on your browser’s extensions page.
2. In Intelligence → Sources, choose Connect browser. Enter its port and one-time code in the extension and allow the local connection. If the permission prompt closes the popup, reopen it.
3. On a page, choose Preview this page. Review or edit the text, confirm, then select Save to Velora.
4. The source appears in this project’s library. Pairing grants permission to import only. Context and training still need separate reviews.

2. Let agents use reviewed project knowledge

1. Open the source and select Propose as project knowledge.
2. In Knowledge, inspect the proposal and choose Approve for context.
3. In Controls, turn on Approved project knowledge.
4. Use Preview context in Knowledge to see which references match a task. References used by recent private tasks appear in Intelligence.

Importing a source alone does not make it trusted knowledge. When retrieval is enabled, selected references can be sent to the provider running your private tasks. Choose an appropriate local provider and data policy in Controls when data must stay on your computer.

For shared team work, use reviewed room knowledge and scoped agent roles. A private project’s sources are not automatically published into a shared room.

3. Turn permitted sources into training examples

1. Open a source in Intelligence, record the rights basis and permission details, then grant training permission.
2. In Controls, enable Dataset curation and local training.
3. In Datasets, choose the evidence revision and write a prompt, reviewed answer, and exact supporting quote.
4. Inspect the examples and screening findings. Approve training use only after reviewing the dataset.
5. Export JSONL and provenance, or use the approved dataset to prepare a local training job in Models.

Velora keeps examples linked to source revisions and checks permissions again at export. Revoking source permission removes dependent approvals. Copies already exported remain under your control and cannot be recalled by the app.

4. Train an open-weight model you own

1. In Intelligence → Models, choose Set up under Open-weight models you own. Velora installs Apple’s MLX into a Python environment only it uses (python3 is required; macOS offers to install it the first time).
2. Download a model — Qwen2.5 or Qwen3, Llama 3.2, Gemma 3 or Mistral, 0.5B to 7B — after accepting its license. Each download is pinned to an exact revision and added to the project’s registry.
3. In Datasets, choose Draft examples: Velora turns each section of a source into a question, its answer and an exact quote. Edit or drop any, create the dataset, and approve it for training.
4. Under Train a model on your data, the form opens on your newest model and dataset. Prepare the job, check it, and start it. LoRA training runs on your Mac; your data does not leave it.
5. When the run finishes, Velora scores it: the base model against the trained model on the same examples, with MLX’s test loss — held-out examples first, the ones it never trained on. Lower is better. It measures your data, not a general benchmark; choose Score again at any time.
6. Choose Use with agents. Velora serves it on your computer and it appears as a local model in every agent and chat model picker. Stop it from the same screen.
7. Choose Export model to fuse the training into a standalone model in a folder you own, with a model card recording its base model and license, the dataset, the training settings and the scorecard. The export joins your registry, so you can serve it or train it again on new data; Show exported model opens its folder.
8. To try another base model, choose it under Carry this dataset to another base model, prepare the run, and compare the two scorecards.

Training and serving need an Apple silicon Mac; 7B models need 16 GB of memory or more. On Windows, export approved datasets to train anywhere. Dataset curation and export are available independently of training.

Use Evaluations to compare pinned configurations at the same repository revision. Register and review a serving model in Models, evaluate the candidate, and approve deployment before changing routing. You can roll back to the prior configuration. Training does not guarantee a quality improvement.

5. Train on another computer you own

1. On the computer that has the model, turn on Intelligence → Models → Host training on this computer, or run velora train-host --models ~/models on a Linux box with MLX. It shows its address and a pairing code.
2. On your laptop, open Intelligence → Models → Train on another computer, and pair with that address and code. The host’s models appear in your registry and the training form.
3. Prepare and start a run on a host model as usual. Velora sends the approved examples to the host, which trains and scores there; the run and its scorecard appear here. Use with agents serves it on the host, and it appears in every model picker; Export makes it one of the host’s models.

Traffic between the two computers is encrypted with a key derived from the pairing code, which is never sent over the network, so it works on office Wi‑Fi without certificates. Anyone on your network with the code can train and chat with the host’s models: keep it like a password, and make a new one to unpair everyone. Local only projects can’t train on or use another computer. A model on a computer you paired sees personal details but not detected credentials.

6. Keep or take out personal details

When you create a dataset from drafted examples, and when you prepare a run, Velora shows how many personal details the examples mention — email addresses, phone numbers, your private terms — as counts, never values. Keep them, because the model is yours, or take them out first; they are replaced with labels such as [email] and [phone]. Detected credentials always come out. A finished run that kept personal details offers Retrain without personal details, and an exported model’s card records which choice was made. Detection is not exhaustive.

7. Learn from your day

1. In Sources, turn on Learn from your day → Apps and windows. Every 20 seconds Velora notes the app in front and its window title and keeps a journal for each day. On a Mac, macOS asks once before Velora can read window titles.
2. Add folders: text documents in them — notes, docs, code — are imported as Sources, and again when they change. Hidden files and dependency folders are skipped.
3. Save today’s journal to Sources whenever you like; yesterday’s is saved when the day changes. Review a journal before approving it for context or training: window titles contain whatever an app puts in them.

It never records keystrokes or the screen, and skips password managers, private browsing windows, anything you add under Never record, and time away from the computer. Pause it for an hour, or choose Delete captured activity to remove every recorded day, journal and imported file — and anything built on them.

What stays with you

Evidence and project intelligence are stored locally. Knowledge retrieval, outcome capture, adaptive routing, and training start off. You decide what becomes context, what becomes a dataset, when training runs, and when a model changes. Export project intelligence in Controls, and export approved training data in Datasets.

Download Velora · Connect agents and models · Release notes