AI chat for your website
Sensever puts an advanced chat on your website with a built-in AI assistant. The assistant answers visitors around the clock from your own materials, and your team sees every conversation and can join at any moment. This guide covers why it helps, how to set it up step by step, how the assistant learns, what data it sends to the model and what you can configure.
Why an AI chat on your site
Answers at any hour. Visitors get a reply in seconds at night, on weekends and during peak hours.
Less routine for your team. Typical questions about prices, delivery, opening hours or how a feature works are answered from your materials, and agents handle the cases that need a person.
A person is always one step away. The assistant works inside an ordinary request. An agent can read along, take the request or reply, and the assistant steps back.
One inbox. The same assistant can answer requests from the widget, the contact page and connected messengers: WhatsApp, Telegram, MAX and Instagram.
Set it up step by step
Create a project. In Projects choose "New project", enter its name and your site's domain, for example
example.com, and click "Create project".Install the widget. In the project, open "Installation", click "Copy code" and paste the code into your site before the closing
</body>tag. Each site of the project has its own code; add more with "Add site". The site must open over HTTPS.Check the chat. Open your site, click the chat button and send a test message. It appears in "Requests" of this project.
Open the Assistant section. In the project, open the "Assistant" tile. Under "Who answers customers" choose Sensever model.
Choose the bot. "Project bot" is created automatically and gets the project's name. You can change its name in "Bot name" and pick a "Personality". Save the settings: the assistant can stay off for now.
Give it knowledge. Tick your channels under "Knowledge base channels" and click "Save", and add your own documents and tables under "Files and tables". The next section explains both.
Test it. Ask a few real customer questions in the "Test chat" and look at the answers and their sources.
Turn it on. Switch on "Virtual assistant" and save. From now on it answers new text messages from customers in the project's requests.
No website at all? The same assistant answers in WhatsApp, Telegram, Instagram and MAX: create the project without a website and connect the messengers. Details: Project AI assistant: setup, knowledge and hand-over to your team.
The assistant can be turned on only when it has something to answer from: at least one knowledge base channel, one processed file or one enabled data connector method.
For developers: the widget installation guide covers language, appearance, your own button and JavaScript events.
How the assistant learns
The assistant doesn't need training in the machine learning sense. It looks up answers in the materials you connect, every time a question comes in. When you update a material, the next answer uses the new version.
Knowledge base channels. Write articles in Studio and publish them into your channel: instructions, terms of delivery, FAQ, product descriptions. Under "Knowledge base channels" the project lists the channels you own or manage. A ticked channel gets the project bot as a read-only member, and its published articles become the assistant's knowledge. The bot doesn't post, invite anyone or change anything in the channel. The assistant reads an article by its headings and quotes only the sections it read, so clear H2 headings make answers better.
Your own files and tables. Under "Files and tables" click "Add file" and upload a file from your device or choose one from Storage.
Documents: PDF, DOCX, TXT, MD and HTML, up to 25 MB each. The text is split into sections by headings, and the assistant finds the relevant fragments by words and by meaning, in any language.
Tables: XLSX and CSV, up to 20 MB, such as price lists, stock levels or a list of branches. The first row is the header. In "Columns" you can rename columns and set their types (text, number, date, yes/no): the assistant then filters, sorts and sums by them, for example "the cheapest model in stock".
Each file shows its state: uploading, waiting for processing, processing with progress, ready, or an error with the reason. "Replace" uploads a new version of the same file, and until it is processed the assistant keeps answering from the previous one. "Remove" takes the file out of the knowledge base, while the file itself stays in Storage. Files use the project owner's Storage space. A project can have up to 100 files; a PDF is read up to 500 pages, a table up to 50,000 rows and 60 columns. A scan without a text layer can't be read yet.
Test chat. Ask the way a customer would. You see the answer, the sources it used (file, section, pages or table rows), which searches ran, how many model calls and tokens it took, exactly what the customer would see, and whether the conversation would be handed over to an agent. Nothing is sent to anyone; tokens are counted as usual.
What the assistant sends to the model and what it doesn't
The answers are written by a language model of xAI (Grok), which we call through its API. For each answer the model receives:
the instructions: the bot's name, the project's name, the personality and the answering rules;
the list of knowledge base channels and their section titles;
the customer's current message and the last four text messages of the same request, which can include your agents' and the assistant's replies;
the fragments the assistant found and read for this question: sections of articles, fragments of your documents, and for tables the column names with a few sample values plus the rows a query returned (at most 20);
for data connectors: the description of each enabled method and the parameters it asks the customer for, and after a call only the response fields you chose. Your API keys, fixed values and the customer details a method takes from the request are not sent to the model.
What is not sent to the model:
your files themselves or links to them: only the fragments found for the question;
the customer's contact details, account data, visit history or other requests: only the message texts listed above;
photos and files the customer attached;
anything the customer can't access: the assistant reads only what is open both to the bot and to the person asking, and when several people are in the conversation, only what is open to all of them;
secret keys of common formats, such as
sk-…keys and Bearer tokens: they are replaced with a placeholder before sending.
Keep in mind that whatever a customer types in a message goes to the model as part of the question.
Searching your documents by meaning uses vector representations (embeddings), and they are computed on Sensever's own servers, as is text extraction from PDF files: your documents are not sent to a third-party embedding service.
Model usage is counted per project: under "Usage over 30 days" you see input and output tokens by model, and the test chat shows them for each question.
Handing over to your team
The assistant answers only from your materials and doesn't invent features, prices or actions. When the answer isn't there, it says so instead of guessing. If the project has its own files or data connectors, the assistant tells the customer that it is handing the question to an operator, and the request goes to your team as unanswered. A customer can also simply ask for a person. Messages without text, such as a photo on its own, go to your team.
When an agent takes a request (an assignment or their first reply), the assistant is paused in it; the ⏸ / ▶ AI button in the request header brings it back without leaving the chat. Agents can open a conversation to read what the assistant said without taking it over. If the model fails, the customer gets a short notice and the request goes into the normal routing. Details: Customer support: requests, team and routing.
What you can configure
Bot name. By default it is the project's name; customers see it on the assistant's messages.
Personality. Eva is warm and attentive, Leo is calm and concise, Leia is lively and friendly. The personality changes only the manner of replies, not the facts. The name and personality belong to the bot and are the same in every project where it is selected.
Knowledge. Which channels are ticked and which files and tables are added.
Source names. "Show customers the names of source files under answers" adds a line with the file names under each answer. It is off by default; links to the files are never shown.
Data connectors. Live data from your own system through its API, read only: see Data connectors: the assistant reads live data from your API.
AI in each channel. "AI answers in this channel" turns the assistant off for a single messenger, the widget or the contact page: see Channel access: who handles requests from each channel.
On and off. The "Virtual assistant" switch. Knowledge and the test chat work while it is off, so you can prepare everything first.
The assistant replies in the language of the customer's last message, usually in two or three short sentences, and links only to the articles it actually read.
Your own bot instead of our model
If you want the replies to come from your own code, choose Your server under "Who answers customers" and select the "Bot that receives messages". Create the bot in chats with "Create → Bot" and add a webhook URL in its settings: bots without a webhook can't be selected. In this mode Sensever only forwards customer messages to your bot's webhook and passes your server's replies back; we generate nothing ourselves, and the model, the data and the logic are yours.
Read more: How a bot answers website visitors as a project and Creating and managing bots.
Live data from your systems
Data connectors let the assistant answer questions about a customer's own case, such as an order status or a booking, with live data from your system's API. You add a connection once (the base address and the key), describe its methods, check each one, and the assistant calls them when a question matches. Connectors only read data, the model never sees your key, and only public HTTPS addresses are allowed. Step by step: Data connectors: the assistant reads live data from your API. All the assistant's settings in one place: Project AI assistant: setup, knowledge and hand-over to your team.
See also: Sensever: what it is and where to start and Website widgets: support, comments and private access.
Bình luận