kovacCustomers ask the same questions your help centre already answers. A chat widget can answer them at any hour, but only if it answers from your content and says where the answer came from, rather than producing something plausible that your support team then has to correct.
Grounded in your help content
You point kovac at the material your support team already maintains: help-centre pages, product documentation, policy documents, FAQs. A website source crawls your public pages on a schedule, including pages built with JavaScript, so the bot learns about a change soon after you publish it. Files and other sources can sit alongside it in the same collection.
Search combines keyword and semantic matching and reranks the results before the agent writes its answer. Every answer carries citations that open the passage it came from, so a customer can read the original and your team can see exactly why the bot said what it did.
One script tag on your site
The widget is added to any page with a single script tag. It is backed by one of your agents, and you style it from a fixed set of theme settings so it matches your site. There is nothing for your visitors to sign up for.
The widget carries a publishable key that can start a chat session and nothing else. It cannot read your documents, change your settings or reach any other part of your account, so having it in your page source is safe.
Limits that apply per visitor
Each visitor receives a signed, anonymous handle, which lets kovac remember their conversation when they reload the page and apply rate limits per person rather than per network address. A single visitor cannot exhaust your capacity, and a whole office behind one address is not blocked because of one heavy user.
Guardrails on the way in and out
Checks for jailbreak attempts and unwanted content run on the customer’s question and on the agent’s answer. You decide what the agent is for in its instructions, and you choose which model it uses from the models your workspace has configured.
Improve it with evidence
kovac can generate test questions from your help content for your team to review. You then measure how often the right passage is retrieved, compare settings against each other, and see whether a change helped before customers see it. Conversations are traced, so when an answer goes wrong you can see what was retrieved and why.
Usage is metered per workspace, and quotas cap storage and ingestion, so the cost of the bot is visible and bounded.