Training and bonus tokens

Train Oasis from assistant replies to improve the product and earn bonus AI tokens for the day when your submission qualifies.

What this is

Training is feedback you submit from an assistant message using thumbs up or thumbs down, a category, at least one badge, and a written comment (30+ characters). It is the main "train the browser" loop: you correct a specific reply and explain why, not just whether you liked it. For whether feedback stays on your device, see Does data leave the device?

A privacy-first browser you can train

Oasis is built around explicit training: you choose when to submit structured feedback on a reply—not passive browsing signals labeled as learning. That is different from many setups where a single shared browser profile or account-wide sync spreads bookmarks, history, and context across work and personal use without clear boundaries.

Browser profiles let you separate environments (for example work vs personal) so sensitive browsing data does not mix by default. For customer-success, research, or writing workflows, training helps improve how Oasis handles context-heavy tasks; choose Anonymous when you want to contribute product signal without linking that submission to your account for personalization.

Does data leave the device? (anonymous training)

Anonymous training is not local-only. When you submit training feedback, structured data is sent to Kahana's servers. Anonymous means your Kahana user ID is not attached to that training record—not that nothing leaves your device.

Is the assistant model on-device only?

No. Assistant chat uses cloud model processing for prompts and replies (see Assistant and cloud data). Training is a separate, voluntary upload when you press Submit on feedback—it is not silent on-device model fine-tuning from thumbs alone.

Some builds include optional on-device embeddings for limited semantic features without calling a remote model hub. That local pipeline is separate from training submissions and from full assistant chat.

What gets sent when I train anonymously?

On submit, Oasis sends a training payload that typically includes:

• user_prompt and assistant_reply (the exchange you are rating)
• badges, comment, category, and sentiment (thumbs)
• training_mode: "anonymous" (no user ID on the training record)
• Optional include_context when enabled in the UI—limited session or tab context may be attached

See Example training payloads (JSON) for the shape.

What is not sent with training?

• Your saved-password vault or imported logins
• Bulk bookmark or browsing-history databases
• On-device semantic embedding indexes (when using the local pipeline)

Do not paste passwords into the assistant or training comment fields. Boundaries: Import and assistant boundary, Interaction data.

Retention and deletion practices are described in our privacy policy; contact support if you have account-specific questions.

Why train

When a reply worked well, your input helps Oasis stay accurate and helpful. When something missed the mark, showing us what went wrong teaches the assistant to do better next time.

Personalized vs anonymous

Personalized links training to your account so Oasis can tailor responses and improvements for you.

Anonymous does not link your user ID on the training record and helps improve Oasis for everyone. Prompt and reply text are still sent to Kahana and stored in both cases—anonymous is not on-device-only. See Does data leave the device? and Train anonymously: persistence, sessions, and logins.

If you are signed in, qualifying submissions in either mode can earn bonus tokens for your account for that UTC day.

Train anonymously: persistence, sessions, and logins

"Train anonymously" is a per-submit choice on an assistant reply—not a temporary mode that resets when you close Oasis, and not the same as keeping feedback only on your device (see Does data leave the device?).

Does anonymous training persist across sessions?

Yes. Submitted training is saved on Kahana's servers. Anonymous means your user ID is not attached to that training record, so it is used for aggregate product improvement, not to build a profile tied to your account.

• What persists: The submission itself (badges, comment, and the prompt/reply text we store under our privacy policy).
• What does not persist (by design): Account-linked personalization from that submission. Choose Personalized if you want feedback to help shape your experience over time.
• Per submit: You pick Anonymous or Personalized for each training submission; it is not "anonymous until you restart the browser."

You must still be signed in to Oasis to earn bonus tokens. Signing in does not convert an anonymous submission into a personalized one—the training_mode flag on the payload is what matters.

How is this different from "Chrome with a chatbot"?

Many AI browsers are a standard Chromium shell plus a chat sidebar. Oasis training is structured feedback (thumbs, badges, and at least 30 characters in your own words) with an explicit anonymous vs personalized choice—not passive browsing telemetry labeled as training.

What about sites that require login?

Training applies to assistant interactions (your prompt and the assistant reply), not to harvesting your site passwords or exporting your logins into a training profile.

• Browsing: You can log into websites in Oasis normally; those sessions use standard browser cookies and site storage.
• Anonymous training: Even when optional context from a logged-in tab is included (include_context), choosing Anonymous still means the training record is stored without your Kahana user ID.
• What we do not do: Use anonymous training to deanonymize you on third-party sites or to collect site credentials for Kahana.
• Password vault: Training does not include your saved-password vault. Imported logins stay in the on-device login manager; see import and assistant boundaries.

Bonus tokens

Qualifying submissions earn +1,000 bonus tokens for today (UTC) in either mode. Unused bonus allowance does not roll over. You can train as much as you want: each qualifying submission earns today's bonus.

Bonus tokens increase your effective daily allowance for assistant use. They are not cash and do not carry to the next calendar day.

What counts as qualifying

You must be signed in. Your submission must include:

• A chosen category
• At least one badge
• A comment of at least 30 characters

The assistant checks these rules before submit. The server applies the same rules when granting bonus tokens. See Example training payloads (JSON) for the shape, including a workflow-style comment example.

How to use it

• Open training from the thumbs up or thumbs down control on an assistant reply.
• Choose categories and badges that match what happened.
• Write at least 30 characters explaining what worked or what to improve.
• Pick Personalized or Anonymous, then submit.
• Check the daily token bar after a qualifying submit to see bonus tokens applied for today (UTC).

Beyond thumbs: training the browser for real work

For writing, research, customer, and context-heavy work, useful corrections are often more specific than thumbs up or down alone—for example:

• "This source matters—prefer it when it is open."
• "Ignore this tab pattern / do not summarize background tabs."
• "Summarize in my format (bullets, tone, length)."
• "Don't interrupt during research mode."

Oasis v1 captures that nuance primarily through your category and comment, not through thumbs alone. Badges such as Helpful or Fast describe how this answer felt; your comment carries the reason and any rule you want remembered.

Why the required comment matters

A pure badge (or thumbs only) records a reaction. A required free-text comment records your reason—what to repeat, stop, or change. That is especially important when the same thumbs up could mean "good recipe once" or "always use this citation style."

Tip: In the comment, say whether feedback is for this task only or ongoing behavior (see preference scope below).

Badges vs preferences you want remembered

In the chocolate-cake example, Helpful and Fast tell you the reply worked well. They do not by themselves say whether Oasis should remember:

• a source preference,
• a formatting preference,
• a tolerance for detail, or
• that only this one answer was good.

Put that distinction in your comment and pick a category that matches (for example answer quality vs behavior preference). See the two example payloads below.

Preference type and scope (today and next)

Today: The submission JSON includes category, badges, comment, and related fields—there is not yet a separate preference_type enum in the payload. Use the comment to spell out intent, for example: "For all future sessions, always…" vs "Only for this project today…."

Next layer (product direction): A dedicated preference type (source, format, interruption, research mode, etc.) plus scope (global vs this task) would make it easier to apply feedback correctly in browser workflows. Until then, treat the comment as the carrier for that detail.

Inspectable, editable preferences—not silent learning

For trust—especially in writing and research—you should know what Oasis is learning from you. Training submissions are explicit (you choose category, badges, and comment and submit). They are not a hidden background scrape of browsing.

• Personalized training links feedback to your account so improvements can apply to your experience over time.
• Anonymous training improves the product in aggregate without tying the record to your user ID.
• A future preference library (review and edit rules you have taught) is the direction; today, your comment text is the clearest record of what you meant.

If you need a hard rule to apply on every reply, state it plainly in the comment until preference settings are inspectable in the product UI.