Sensever virtual assistant: answers, sources and teamwork
The Sensever virtual assistant helps you use the service based on the official channel content. It isn't an agent with access to every account, and it doesn't perform actions on your behalf. This article explains how to ask a question, how sources are chosen, when the team joins in and what the model usage figures mean.
What the Sensever virtual assistant can do
The assistant helps with chats, finding people, contacts, channels, articles, Flow, projects, comments and the API. It works in official support at sensever.com/help and in the same request inside chat. It answers briefly, based on the published content of the Sensever channel.
It doesn't manage your account, change tasks, send invitations or search for people in private user conversations. If you ask it to find someone, the assistant explains where search is, but doesn't run the search on your behalf.
How to ask questions and continue the conversation
You can write in plain words: "How do I find a user?", "Where do I add a contact?", "How do I message someone first?", "Why was this request assigned to me?". You don't need to add the word Sensever to every question. If a phrase could mean different things, the assistant clarifies your intent: for example, whether you mean a user of the service or a support agent.
The answer takes into account the last four text messages, and the preliminary topic check uses a shorter context. A follow-up like "are you sure?" refers to the previous substantive question. In a long conversation, repeat the feature name and the action you need: earlier details may fall out of the context.
Where answers come from and how tokens are saved
Every message first goes through a short topic check. A greeting, a thank-you or a question about the assistant itself is handled by that same check, without reading the knowledge base. For a real question, the assistant gets the structure of available topics and chooses an article, its headings and the relevant sections. The full text of every article isn't sent with each question.
Links in answers should point to the sources that were read. You can read the knowledge base articles yourself in the official Sensever channel. Updating a published article updates the available source; you don't need to create a separate copy of the same article for the bot.
Why an answer may be wrong or missing
Possible causes include a misclassified topic, no clear instructions in the knowledge base, an unavailable source, a request limit or a temporary model failure. Replying "this question isn't about Sensever" to a question about a feature of the service is an error, not proof that the feature doesn't exist.
Write the name of the section and the action you want to perform. If a previous answer was wrong, you can point out the mistake directly and rephrase the question. The assistant should check the source again rather than treat its previous answer as confirmation.
How a support agent joins
Write in the same request that you need help from an agent and describe the problem. The team sees the request history. The assistant must not claim that it called a person or assigned someone if that didn't actually happen.
While the AI is handling a conversation, regular auto-assignment doesn't take it away from the assistant. An agent can open the conversation to view it; when an agent takes the request or replies, automatic answers stop. If the AI fails, the request can go into regular distribution.
AI and the widget on your site
Any project can have its own AI assistant on its website: open the project's "Assistant" section, choose "Sensever model", tick knowledge base channels, add your own documents and tables and turn on "Virtual assistant". The whole setup, what the assistant sends to the model and how it hands over to your team are in AI chat for your website.
You can use the regular support widget and comments within their available settings. For a closed site, access to a private discussion is set up through the site's server-side integration; a hidden page doesn't by itself replace access checks.
Bots: who answers and what they can access
The one answering isn't a "project assistant" but a bot — an account just like a user's, except that it belongs to a person. You create a bot in chats: Create → Bot. It has its own name, its own personality and its own way of working: answer with our model, or receive messages on your webhook and reply with your own code.
What a bot knows depends on where it has been invited. A bot is added to a channel as a regular member with the "Reader" role, and that channel's content becomes its knowledge. Remove it from the channel and the content stops being used immediately, with no changes to the project settings. A project's assistant also answers from the files and tables added to that project under "Files and tables".
A project only selects and enables a bot. That's why the same bot selected in two projects remains a single assistant: its name and personality belong to it, not to the project.
An answer is based on the intersection of permissions:
What the bot can access — the channels it's a member of.
What the person asking can access — private content won't make it into an answer for someone who doesn't have access to it, even if the bot can read it.
What everyone in this place can access — in a group or channel, the assistant doesn't say out loud anything that's closed to anyone present.
That's why a single bot can easily serve both customers and employees: a customer gets an answer based on public content, while an employee also gets answers from private content they have access to.
AI helper: sources and statistics
In a project's "Assistant" section you choose knowledge base channels and add files and tables. Separately, an API key with the permissions you need lets your own code read the same channels through the library API: the tree, headings, texts and links to sources. The official Sensever channel uses this same library.
What is counted
The project's "Assistant" section shows the model usage over 30 days. For a connected assistant, the actual input, output and cached tokens of every model call are counted, including the question's topic check. The cost is shown based on the provider's data if the provider returns it. Recording the same request twice doesn't increase usage.
How the official assistant works
The Sensever virtual assistant answers in the widget at /help and in the same request in chat. For questions about the service, it first checks the topic, then selects an article, a heading and the relevant paragraphs of the official knowledge base. It gives a short answer and a link to the source. Off-topic questions don't get the knowledge base context. It doesn't read personal tasks or other people's discussions. When an agent takes the request, auto-replies turn off.
API and AI: what's connected, sources and tokens
The unified data API works independently of AI. The official Sensever assistant is enabled at /help and reads the public knowledge base of the official channel. Projects connect the assistant in their "Assistant" section; a universal /api/v1/assistant/answer endpoint isn't available yet.
Project sources
Under "Knowledge base channels" in the project's "Assistant" section, choose channels you own or manage. This choice does two things: it limits the sources of the project key and gives the assistant bot read-only membership in those channels. The bot doesn't publish, doesn't invite people and doesn't change anything in the channel.
Usage tracking
The project statistics show API requests and errors, and for a connected model, the actual input, output and cached input tokens. Classification and source-reading steps are counted too. The cost is shown based on the provider's data when available; these are usage statistics, not a promise of connected billing charges.
Saving tokens
Greetings don't need the library to be sent. For a question, the topic is determined first, then the relevant content and section are selected. The public assistant doesn't get the private knowledge base, even for a signed-in visitor. The full conversation history and all articles aren't sent with every request.
Komen