What Onflow AI knows
Onflow AI is the assistant behind the chat in the help widget. It is not a general chatbot with an opinion about Onflow Ads. Every answer it writes is built from what the platform itself says: a short product brief that rides every question, plus the Onflow Ads documents that are published right now, retrieved fresh for your particular question.
That is the whole design, and it is why the answers are usable: it knows the product it supports, it quotes what the platform actually says, and when neither covers your question it tells you so instead of guessing.
The product brief​
Whatever you ask, the assistant starts from a compact description of Onflow Ads itself — what the platform is, the five things an account can do, the order the first four steps go in, the connection gate, the ways to sign in, and the shape of the wallet, the plans and the bot. When your question is about money or plans, the live figures ride with it: every plan price, the marketplace fee model, the top-up fee ladder and the withdrawal fee ladder, read straight out of the configuration that charges them rather than typed into a crib sheet that could go stale.
This is why an ordinary question — how do I start using this? — is answered in full even when nothing in the corpus matches the words you used. The brief is the floor. It is not a substitute for the documents: where a published page and the brief ever disagree, the page is newer and wins.
The four retrieved sources​
| Source | What it covers | How it gets in |
|---|---|---|
| This guide | Every page of docs.onflowads.com — plans, wallet, boost, marketplace, cross-promotion, subscriber exchange, the bot, support. | Enumerated from the guide's own sitemap, so a new page joins the corpus by existing. |
| The three policy documents | The published Terms, Privacy Policy and Refund Policy — including the extra clauses the team publishes onto each of those pages. | Each is fetched as one long document, clauses and all. |
| Knowledge-base documents | Product documents the team has uploaded — the text is extracted and chunked page by page. | Added from the support desk; no code change, no redeploy. |
| The help-centre FAQ | Hand-written question-and-answer pairs, each with a deep link to the page that goes deeper. | Written by the team and matched with the highest precision of the four. |
Everything above is content you can read yourself, and so is the product brief — it says nothing about Onflow Ads that this guide does not, and its numbers come from the same configuration the checkout charges you from. There is no internal wiki, no private policy and no separate set of rules the assistant answers from. If Onflow AI knows it, you can go and read it — and the buttons under the answer take you straight there.
It stays current on its own​
The corpus is not a snapshot taken on release day.
- A background pass re-reads every source on a rolling cadence, roughly every few hours, so an edited policy clause or a rewritten guide page reaches the assistant on its own.
- Only the passages that actually changed are re-processed, so the sweep is cheap and can run often.
- Sections that no longer exist are dropped, once the pass has genuinely reached the site — an outage never empties a good corpus.
The practical consequence: if a page in this guide says something today, the chat will say the same thing today.
How an answer is built​
- Your question is matched by meaning, not by keyword. You do not have to guess the platform's vocabulary — ask the way you would ask a person.
- The closest passages are pulled from both corpora: the hand-written FAQ pairs first, then up to six sections of the guide and the policies. If nothing clears the bar, a second, looser pass runs and whatever it finds is offered to the assistant as possibly related rather than as an answer.
- The answer is written from the product brief and those passages, together with what has already been said in this conversation — roughly the last twenty exchanges ride along, so it remembers what it told you. Nothing else counts as knowledge: no figure, screen, button, policy or timeline that is not in front of it may appear in an answer.
- The links are attached — built from the passages that were actually retrieved, and only from the confident ones. An answer that leans on the brief, or on a loose match, arrives with no buttons rather than with a button that points somewhere it did not come from.
How the answers read​
The assistant is told to be complete rather than brief, and the style follows from that:
- Plain text, never markdown syntax. No asterisks, no hashes, no bracket-and-parenthesis links.
- Bullets for anything enumerable — steps, methods and options each go on their own line starting with a
•. - Usually under 120 words, longer only for genuine step-by-step instructions.
- Every way, not the first way. When the sources describe more than one route to the thing you asked — sign-in methods, payment methods, contact routes, plan options — it is required to name all of them. A visitor told about one method when three exist was misinformed, not helped.
The read-it-yourself buttons​
Under an answer you will usually see up to three buttons carrying a document icon and a page name. They are the receipts.
| Property | Detail |
|---|---|
| Where they come from | The passages that were actually retrieved for your question — built server-side, never written by the model. |
| Where they can point | Only onflowads.com or docs.onflowads.com. Nothing else can become a button. |
| How precise they are | A guide link carries the section anchor, so it opens on the exact heading that answered you. |
| How many | Three at most. Hand-written FAQ links first, then the distinct guide and policy pages. |
| How long they last | They are stored with the answer, so they survive a refresh and are still there when you reopen the conversation. |
The button usually holds the fuller version of the answer — the caveats, the edge cases and the numbers. On a metered plan it also costs you nothing, while a follow-up question costs two messages. See Your allowance.
It will not invent an answer​
This is worth being blunt about, because it is the difference between an assistant you can act on and one you cannot.
- When the sources do not cover your question, it says so. It does not fill the gap with something plausible.
- It never promises refunds, credits, timelines or policy exceptions. The only commitment it can make is one the published policy already makes.
- It never says it will connect, transfer or escalate you, and it never asks for your email address. It cannot do either. When a person is needed it says so plainly, and the handover question is asked by the system itself, in the same message, after the answer. See Asking for a person.
- It does not paste URLs into its prose. The relevant pages arrive as buttons underneath instead.
Nothing Onflow AI writes commits Onflow Ads beyond what the published policies already commit to. If it tells you a policy applies to your case, the team's review is the one that counts.
What it is not grounded in​
- Your account — signed out it can see nothing about it. Signed in, on every plan, a separate read-only layer can look up your own references, reliability history, refund standing and partner suggestions. That is its own page: Answers about your own account.
- Anybody else's account, on any plan, ever.
- Anything unpublished — internal queues, another user's data, decisions not yet made.
When it cannot answer​
It should be rare, because the brief covers how the product works and the corpus covers what it says. What is genuinely outside its reach is anything that needs somebody to look inside an account — a refund decision, a payout, a charge that did not land, a penalty or an appeal — and anything that has nothing to do with Onflow Ads, which gets one friendly line saying so.
Two different things can then happen, and neither is a dead end.
A person is needed. It says so, and the system offers to fetch one in the same message. See Asking for a person.
The assistant itself is briefly unavailable. Your message still goes somewhere useful:
- A confident match in the help centre — up to three answer cards, best first, each with the question in bold, the answer word for word from the corpus, and a Learn more → link where a longer page exists. Still need help? Email the team sits underneath and carries your question into the message form.
- No confident match — the message form opens with your question already typed into it, under the banner "Onflow AI is temporarily unavailable — your message will reach the team by email instead."
Either way, your question reaches somebody without you retyping it.
If something goes wrong​
| What you see | What it means | What to do |
|---|---|---|
| "I'm having trouble reaching my knowledge base right now." | The assistant could not retrieve or generate this turn. The same message offers a human | Reply yes to be handed over, or try again in a minute |
| An answer with no buttons under it | Nothing in the corpus matched confidently enough to cite | Take the answer with the caution it deserves, and ask for a person if it matters |
| An answer that says a person is needed | The question needs somebody to look inside your account, or to make a decision | Say "talk to a human" — or reply yes to the offer — and the whole conversation goes with you |
| Answer cards instead of a conversation | The assistant is briefly unavailable and the help centre answered | Read the cards, or press Still need help? Email the team |
| An answer you can tell is wrong | It missed | Say so in the next message; it can correct course. If it keeps missing, ask for a person |
Related​
- Chatting with Onflow AI — the conversation itself, the allowance and the handover.
- Answers about your own account — what it can look up once it knows it is you.
- The instant first reply — the same grounding, on a ticket you opened from the form.
- Opening a support ticket — writing straight to a person instead.