Chat basics¶
What happens after you press Send¶
The answer is written into the chat in front of you, word by word, the way a person types. You can read the opening while the rest is still being written, and stop as soon as you have what you needed.
Answers come back as formatted text: headings, lists, tables and blocks of code. Wide content, such as a broad table, scrolls inside its own block instead of stretching the page sideways.
The line that appears before the words do¶
An answer rarely begins the instant you press Send. Until the first words arrive, one line with three pulsing dots — the activity line — tells you what is being done right now:
| What you see | What is happening |
|---|---|
| Searching the web for "luxembourg weather" · 4s | The live web is being searched. The words it searched for are shown, and the timer counts on. |
| Looking in your documents | Your own uploaded files are being searched. |
| Running Python | A small program is being run, to work something out or to build a file. |
| Asking Research buddy | One of your agents has been brought in by name. |
| Using Google Drive | A connected system is being consulted. |
The moment text starts to flow, that line disappears — the answer itself has become the sign of progress.
The line also reports how a step ended before moving on. A finished search reads Searching the web · Found 5 sources. Something that did not work reads Didn't work; the assistant will usually carry on without it and tell you in the answer what it could not check.
Looking in your documents sometimes appears for a moment and vanishes again on a straightforward question. That is the right outcome rather than a fault: your files were searched, nothing in them was relevant, and the answer was written from what the model already knew.
Every one of these states is written out in words as well as marked by colour, and each change is announced to a screen reader as it happens.
While the answer is being written¶
Stop ends the answer straight away. Whatever had already been written stays in the chat, marked as stopped, and you get that one message — not a second one appearing later.
Stop ends the work, not just the view of it. The model stops writing, any step still running is cancelled, and nothing further is sent to your model supplier on your behalf. What had already been sent before you pressed it has been used and is charged for, and it appears in your organisation's figures marked cancelled — so a stopped answer is cheaper than a finished one, but not free.
You can leave the chat, or refresh the page, and come back: the text produced so far is still on screen, and the timer counts the real elapsed time on the server instead of starting again. A long job keeps counting, and that is how you tell "still working" from "stuck". If you have the same chat open in a second browser tab, the answer arrives in both at the same time, which is intended.
If your organisation's model supplier has a momentary problem, you see the model call did not complete — retrying (attempt 2 of 10) rather than a long silence that ends in a poor answer. A single piece of work will keep trying for up to about ten minutes before it gives up, it shows you where it has got to throughout, and Stop always works.
If an answer disappears¶
An answer is saved only once it is finished. While it is still being written it exists on your screen and nowhere else. Refreshing the page during those seconds keeps the partial answer visible, but it is not yet stored.
So if zGentic is restarted while an answer is in flight — most often because a new version is being put in place — that answer is lost, and the message is marked as failed with this wording:
The backend restarted while this turn was running; it was interrupted.
"The backend" is the part of zGentic that does the work, and a "turn" is one exchange: your message and the answer to it.
This is expected behaviour, and there is nothing to report. Ask your question again and it will be answered normally. Nothing else is affected: every earlier message in the chat, and every file already finished, is stored safely.
The Tools sheet¶
The + button beside the message box opens Tools. Everything in it applies to the chat you are in, and is remembered for that chat.
| Control | What it does |
|---|---|
| Retrieval (RAG) | Answers from your own documents. The assistant searches the files you and your projects have uploaded, uses the passages that match, and shows you which file each part came from. |
| Web search | Lets the assistant search the live web and open pages it finds. See Web search. |
| Thinking | Lets the assistant spend longer working the problem out before it answers. The working-out is shown, folded up, under the answer. Useful for a hard question; it makes an easy question slower and more expensive. |
| Upload a file | Adds a document or a picture to this chat. See Attachments and pictures. |
| Agents | Decides which prepared helpers this chat may call on. See Projects, agents and tasks. |
| Connectors | Decides which other systems this chat may reach — Jira, Confluence, a weather service. See Connectors. |
Retrieval (RAG) is the label on the button. RAG is short for retrieval-augmented generation, which is a technical way of saying: look the answer up in your own documents first, then write it.
A control marked Unavailable has not been switched off by anyone. It means the feature has not been set up for your organisation yet. Web search is the usual example: it needs a search account, and an administrator adds it once (see enabling web search).
Choosing which model answers¶
A model is the engine that writes the answer. Models differ in how fast they are, how much they cost, how well they handle difficult work, and whether they can look at pictures.
The model picker sits at the top of the chat. Your organisation decides which models you can pick from and which one is used when you do not choose. A badge marks the models that can look at pictures.
If the model you picked is temporarily out of service, the next model in your organisation's list is used instead, so your question does not fail. You may notice the answer is written in a slightly different style.
Commands that start with a slash¶
Type / at the start of the message box and a list of commands opens — /memory save …, for
example. Commands are carried out by zGentic itself. No model is involved, so they are instant and
cost nothing. They are described in Memory and commands.
Anything that is not a recognised command is sent as an ordinary message.
Lining up your next question¶
If you start typing while an answer is still being written, Queue holds your message and sends it as soon as the current answer is finished. Use it when you already know what you want to ask next.
Finding, renaming and deleting chats¶
- History, in the left-hand navigation, searches everything you have ever asked.
- The pencil beside a chat's name renames it. The first name is taken from your opening message.
- Deleting a chat removes its messages, the files you uploaded into it, and anything made from them.
The two small numbers near the message box¶
The ↑ and ↓ figures show how much text went into the current answer and how much came back. They are counted in tokens. A token is a chunk of text a little shorter than an average word, and it is the unit models are priced in.
Treat the two numbers as a rough gauge: they tell you whether a conversation is getting long and expensive. Owners see the same activity in money under Billing and usage.