Internal notes stay internal
Content marked internal-only is reachable from your workspace and never from a customer surface. There is no prompt to talk it out of that.
Aqvitus answers from your articles, your past resolutions and the sources you connect — and when it cannot answer, it says so, groups the question with every other customer who asked it, and drafts the article.
Drafted from how your team actually answered. Nothing publishes without a person.
Four stores, one index. The AI searches all of them at once — by wording and by meaning, with newer and higher-priority material weighted up — and reranks what comes back.
Help-center articles and internal-only ones, each with an owner and a freshness date so a stale answer surfaces instead of hiding.
Question-and-answer pairs harvested from real resolutions, approved by a person. When one is as relevant as an article, the learned answer wins.
The built-in knowledge base, a website crawl and file imports, each showing its sync status, how many documents it holds and how much of it is indexed.
Import the conversations you have already resolved. They feed both the answers the AI can find and the test cases a change is measured against.
How many pages and files a workspace can hold depends on your plan — see pricing.
Every escalation and every miss is tagged. Identical questions collapse into one record with the canonical wording, how often it was asked and how many customers it reached.
“Can I change the delivery address after ordering?”
“I don’t have a confirmed answer for that. Let me get somebody who does.”
“Can I change the delivery address after ordering?”
Changing a delivery address after you have ordered
You can change it yourself until the order is picked. After that we can reroute it once it is with the carrier — ask us and we will do it.
Drafted from 9 resolutions · cites 2 of them
Example data. The draft is written from how your team answered those nine customers, not from what a model can invent.
When someone edits a draft the AI wrote, Aqvitus compares the two. If the meaning changed — not just the wording — it proposes a learned answer from your version. A resolution the customer confirmed proposes one too.
Every proposal waits for a person, and you can work through them in a batch. Every correction and every escalation also becomes a test case, so the next change to the AI is measured against the thing it got wrong last time.
How a change is tested before it goes live“Refunds take 5–7 working days.” “Refunds reach your card in 3 working days, or up to 10 if you paid by bank transfer.”
Intent refund_timing · from Priya’s edit · waiting for review
The same edit also becomes an eval case in the candidate suite.
Every search carries who is asking: how far they have verified themselves, and their segment, plan and region. Out-of-scope material is not returned — that is enforced in the search itself, not asked of the AI in a prompt.
Content marked internal-only is reachable from your workspace and never from a customer surface. There is no prompt to talk it out of that.
A learned answer written for enterprise accounts in the EU is not offered to a starter account in the US. The scope is part of the record.
Anything that needs a verified customer is not retrieved for an unverified one. The AI asks them to verify instead of guessing.
Retrieval is one half of the context the AI answers from. The other half is the procedure it is following.
The AI agentIntent volumes and their deltas show you which gaps are costing you the most conversations.
AnalyticsAn answer the AI found can cite where it came from, as a card the customer can open.
The chat widget
Answer what you already know how to answer, and find out what you do not.