Conversation and knowledge
What the chattiness dial actually does, why an unprompted reply is never certain, and how to give the bot your documentation so it answers from your project.
The half of Wardnox that talks. Both are off by default, on every plan: a server that installed an anti-scam bot did not ask for a chatbot.
Conversation
When enabled, Wardnox can answer members in your channels.
It runs strictly behind the security engine and cannot loosen it. A message the engine acted on is never answered, so a scam is removed rather than replied to. That ordering is not configurable, and it is the reason the community half can exist at all.
How much it talks
The chattiness dial governs unprompted replies, and it is worth understanding exactly what it does, because the name oversells it.
- Being mentioned or replied to is always answered, at every setting, including zero. That path ignores everything below.
- For everything else, the dial sets how strong an invitation a message needs before the bot answers, and how often it takes a borderline one.
- Even at maximum, an unprompted reply is never certain. The top of the dial joins roughly one borderline conversation in three. Something has to stop a bot dominating a channel, and this is it.
The dashboard shows the current odds beneath the slider, so you can see what a given setting actually means rather than inferring it from the word.
Two more controls shape the same behaviour. A cooldown it will never speak twice within, in one channel — a hard floor no combination of signals can outvote. And a shortest message worth answering, below which it stays quiet unless addressed directly, because "ok" is not a conversational opening.
There is also a deliberate delay before answering. Replying instantly reads as a machine watching the channel.
Where it stays out
You can name channels it must never speak in. Use this for announcements, moderation channels, and anywhere a bot joining in would be actively unhelpful.
Note that a message in a blocked channel is still scanned. Blocking silences the conversational half only.
Forbidden topics
A list of subjects it will not discuss, whatever it is asked. It refuses rather than improvising, and the refusal itself is checked before it is sent.
Project knowledge
This is the feature that makes the bot worth talking to: give it your documentation, and it answers from your project instead of guessing.
Available on Pro and Scale. Without it the bot can hold a conversation but knows nothing about you specifically.
Adding sources
Paste text directly, or point it at a site and let it crawl. Each source is read, split up, and indexed so the relevant part can be found when a member asks something related.
Your plan sets how many sources you can keep. Pro allows a working set for a project's documentation; Scale allows considerably more.
Core facts
Alongside the crawled sources there is a short block of facts you write yourself: the things that must always be right, such as your official domain, where support happens, and whether staff ever send direct messages.
This is the highest-value thing on the page and the most under-used. A bot that knows your official link is a bot that can be asked "is this the real site?"
Getting a wrong answer fixed
If it answers something incorrectly, the fix is almost never to disable the feature. It is to correct the source, or to state the fact in the core block, which outranks whatever a crawl inferred.
Why it says "I don't know"
The model behind it was chosen on a benchmark for making things up rather than on price. A bot that confidently sends a member to a channel that does not exist is not a slightly worse bot; it is training your community to follow instructions from something that invents them. An honest "I don't know" is the feature working.