Chatting

Working in a project

Every conversation belongs to one of your company’s projects. The project you are working in is the open one in the Projects list of the sidebar, under New chat and the search, and holds its conversations; when you are in several, open another to work in it. A new chat starts in the open project, and you choose among the models it allows - only those: the owners and administrators of your organisation choose, project by project, which providers and models each one may use. New chat asks which project the chat will be in when you belong to several, and a new chat says it above the message field, as in “New chat in Legal”. When you open a chat, its project opens in the sidebar and the chat is highlighted.

In the sidebar, each project wears its own colour and initial, which it keeps, and says how many chats it holds; a folded project also says when it was last used. The chats of an open project hang from a line below it, so you always see which project a chat is in.

Each project has a +, shown when you point at it, that starts a new chat in it. To move a chat to another project, drag it onto that project (on a phone, press and hold the chat first), or open the chat’s … menu and choose Move to the project. The chat is then listed under the other project, and its next message is sent to the models that project allows - a model switched off there can no longer answer it, so choose another.

“No model in … yet”. A new project offers no model until an owner or administrator of the organisation chooses its providers. Until then the chat says so. If you are one of them, choose Choose the models: it opens the catalogue, where you turn providers on for the project (see Switching providers and models). Otherwise, ask one of them.

If you are taken out of a project, its conversations - yours included - are no longer reachable for you. They are not deleted: they come back if you are put in the project again.

Choosing a model

The model that answers your next message is named at the top of the chat, with the provider serving it. Tap it to open Choose a model; it lists only the models your current project may use:

Type in Search models or providers to narrow the list.
Point at the ⓘ icon at the end of a model’s line - or tap it on a phone - to see, before you choose it, where its provider is based, the data centres the model runs in, and the certifications its provider holds, with their state (certified, in progress, expired or only self-declared). A place or a certification nobody has published is said as such, never guessed. The card stays open while the pointer is on it: when a certification shows a link icon, click it to read its proof.
Models under Ready can answer now.
Models under Not connected are listed so you know they exist, but cannot answer from Teloneum yet: Teloneum does not provide access to their provider yet. They are greyed out. A model not available in your project is not in the list at all; the administrators of your organisation can make it available there from the catalogue.

If you administer a project, you can also start from the model catalogue: the Chat button on a model opens a new chat with that model already chosen.

If the model you had chosen can no longer answer, Teloneum uses the first model that can.

Choosing the right weight

A heavier model is not better at everything: it is slower and uses more. To help you choose, each model carries markers, lightest first in Choose a model:

The weight, as a disc that fills its ring the heavier the model is: Feather, Light, Medium, Heavy or Very heavy - what the model costs to use, read from the price its provider publishes or from the range its maker places it in. A sentence says what it is usually for: “The right default: analyse, write, code.” for a Medium model, for example.
The performance, as a gauge, from Basic to Excellent, and the speed, in chevrons (Slow, Normal, Fast): read from public tests, and compared with the other models Teloneum offers.
Efficient: no other model is both lighter and better.
The strengths, when the model is among the best in a field: Strong at code, Strong at reasoning, Strong in French, Reads long documents, Uses tools well, Reads images, Reliable on facts.

Pick a weight, or Efficient only, above the list to narrow it. When unsure, take an Efficient model of the weight you need: Medium suits most requests. Go up a step if the answer falls short, down if it is obvious.

Hover over a marker, or long-press it, to see what it tells: “Speed: Fast, 3 of 3 - how fast the model writes its answer, from public tests.”, for example.

A marker nobody has published for a model is not shown, never guessed.

Switching model in a conversation

You can change the model at any point of a conversation: pick another one, and your next message goes to it, together with everything said so far. Each answer is labelled with the name of the model that wrote it, so you always know who said what. Changing the model while an answer is being written does not change the one writing it: the new one answers your next message.

Asking another model the same question

Under every answer, Ask another model lets you see what another model would have said, with the same conversation before it. Pick a model: it answers the same question, with the same files, and its answer is written in place of the first one.

Once another model has answered, a row of model names shows above the answer. Tap a name to read that model’s answer: it slides right when that name is to the left of the shown model, left otherwise, and the conversation scrolls around the row, which stays still. The first name is the answer the conversation goes on from. Ask as many models as you like: each one is added to the row. The models already asked, and those that cannot read the question’s files, are not offered.

The other answers are there to compare, nothing more: your next message goes to the model chosen at the top of the chat, with the conversation as it was - the models never read the other answers. Each one counts against your usage limits like any answer, and while one is being written you can stop it, or wait for it before sending your next message.

Who processed a message

Under every message you sent, a small shield opens Who processed this message: every model that received it, and every service a web search drawn from it reached. At the top, a world map shows every country and data centre your message went to, all answers together; a place a provider only names vaguely, such as “EU data zone”, is not put on the map. The map stays in view while you scroll through the answers. For each answer - the one the conversation goes on from, then those of the other models you asked - you see:

the model that wrote it, the company that provides it, and the country that company is based in;
the data centres it runs in, as far as the provider publishes them: when a model may run in several, all of them are listed;
the certifications its provider holds, and how far each one is proven.

When the answer searched the web, the list also shows the service that checked the query for personal data before it left, and the search engine that received it. A detail nobody published is said as such, never guessed.

These are the details as they stood when your message was processed: a certification that has expired since still shows as it was then. A message sent before this list existed shows today’s details, and says so.

On a wide screen the list opens over the conversation; on a phone it is a screen of its own, and the arrow brings you back.

Sending a message

Write in the Message box and press Enter or tap the arrow. Shift + Enter starts a new line.

If Teloneum cannot reach the model before it starts answering, nothing is saved and your message is put back in the box, so you can send it again without retyping it.

Copying text

Everything Teloneum shows can be selected and copied: your questions, the answers, the titles of your conversations, project names, settings. Drag the mouse over the text, double-click a word or triple-click a paragraph, then copy with Ctrl + C (Cmd + C on a Mac) or the right-click menu. One selection can run from a question to the end of its answer, and over a whole conversation.

To paste text copied elsewhere, click in the Message box and paste with Ctrl + V (Cmd + V on a Mac).

In text copied across several messages, the paragraphs follow one another without a blank line between them.

Sending an image or a PDF

Some models can look at an image or read a PDF. With one of them chosen, tap the paperclip at the left of the Message box and pick your files: they show above the box, an image as a thumbnail, and the × on each takes it out. A message can go with files alone, or with files and text.

In a browser, you can also paste an image you copied, or a screenshot (Ctrl + V, or Cmd + V on a Mac), or drag files onto the page.

A message takes up to 5 files, each of up to 5 MB.
The paperclip offers only what the chosen model reads: images (PNG, JPEG, GIF, WebP), PDFs, or both. Hover it to see which. With a model that reads text only, it is greyed out.
If you switch to a model that cannot read a file you picked, the file says This model can’t read it and the message cannot be sent until you remove it or choose another model.
The files stay in the conversation: tap an image or a PDF to read it in the document panel. Every later message sends them again to the model, as it does the rest of the conversation, so they keep counting in the context window and in your usage limits.
If you later switch to a model that cannot read them, it is told that a file was there, without seeing it.

If a file is refused, your message and its files are put back in the box.

Answers, as they are written

Answers are streamed: the text appears as the model writes it, formatted - headings, lists, bold, tables, code and links you can tap. At the bottom of the conversation, you follow the answer as it grows; scroll up and what you are reading stays still while the answer goes on below, and when it ends. Until the first words arrive - some models reason for a while before answering - you see Thinking….

An image an answer names by its address is not shown: in its place, a frame gives its description and Image not shown: followed by its address. Nothing is loaded until you tap it; tap it to open the address in your browser. An answer may have been swayed by a mail or a page the model read, and loading the image would send its address - and whatever it holds - without you deciding to. Images Teloneum makes itself, such as a chart, show as usual under the answer.

A conversation takes one message at a time: wait for the answer, or stop it, before sending the next one.

Answers from the web

Some models can search the web while they answer: for today’s weather, the news, a price, opening hours, or a fact they want to check. You do not need to ask for it - just ask your question. The model writes its own search, reads the pages it finds, and answers with links to the ones it relied on.

While the model reasons and searches, a single closed Thinking… line stands above the answer, with the seconds it has been at it. Once the model is done searching and writes its answer, it becomes Chain of thought, with how long all of it took: tap it to see every step - each thought, then the searches it led to, Searched the web for “…”, each with how long it took. Tap the number of pages beside a search to see them, and tap one to open it. Some models show their thoughts word for word, others a summary of them; and a model may not think at all before a simple question.

The model knows today’s date. When the answer depends on where you are, say it - or it will ask you.
Its search is sent to an outside search service. Before it leaves, Teloneum checks it for personal data: a person’s name, health information, contact details, an identifier such as a date of birth or an account number, or other sensitive data. When it finds some, the search waits for you: Allow this web search? names the data and shows the search. Choose Allow search to let it go, or Don’t search: the model then answers without it, or searches again without those data. A search you leave unanswered for a while is not made.
Under every search checked, open the chain of thought to read what was found: Personal data checked: no name, age, place, health topic, contact details, identifier or sensitive topic, or what it found - a place, a person’s age, a health topic - and whether that makes personal data. That is how you see why a search waits for you, even one that carries nothing on its own. Not checked means the check did not answer, and the search was put to you all the same.
The searches of one answer are checked together: an age in one and a town in another can identify someone, and the request then says so. When several searches wait at once, one request lists them all - Allow these searches? - and your choice goes for all of them. Once you have allowed some data, the model’s next searches for the same answer do not ask again.
Each search counts against your usage limits, like the answer itself. An answer that searched costs more than one that did not.
Models that cannot search say so when you ask for something recent.

Exact figures and calculations

When an answer needs exact figures - a total, a percentage, the number of days between two dates, statistics on numbers you give - models that can use tools work them out by running a short program, rather than estimating them. You do not need to ask for it. Each program appears as its own step in Chain of thought - Ran code - and tap it to see the program it ran.

The program runs apart from everything else: it cannot reach the internet, your other conversations or anyone else’s.
Within one answer, each step can build on the previous one. A new message starts afresh for now.

Files the assistant makes

Ask for a chart, a spreadsheet, a PDF, a Word document or a presentation and the assistant can make it: the files appear under its answer.

A chart shows right there. A document shows its beginning in a card - a PDF’s or a presentation’s first page, a letter’s first lines, a table’s first rows - with its name above it.
Tap the card, or its open button, to read the whole file in a panel: beside the conversation on a wide screen, in its place on a phone, where going back closes it. Pages scroll, and zoom with a pinch.
Word, Excel and PowerPoint files are shown as PDFs, so they may look slightly different from the original. The download button always gives you the original file.
Saving a file is in the web app only, for now.

How full the conversation is

A model reads a limited amount of text at once: its context window. Every message you send goes with the whole conversation so far, so a long conversation fills it up.

The small ring beside the send button holds a smaller green ring that fills up with the conversation. Tap it: Context window, at the top of the panel it opens, shows how much of it the conversation takes, for the model of the conversation - the last one that answered in it - for example 24.9k / 270.0k (9%), in tokens, the pieces of text a model reads (a word is often one or two). Your usage limits follow below it. The bar is green, and turns orange past 80 %. When it is nearly full, start a New chat: a model may refuse a conversation longer than its window. Switching to a model with a larger window also gives you room.

The bar appears once a conversation has messages, for models whose window is known. It counts everything your next message sends, the instructions Teloneum gives the model included, as the model’s provider counted it. How much a model thought before answering is not counted, nor what a web search or a code run gave it: neither is sent again with your next message. The answers of other models, asked to compare, are not counted either.

The second figure is the model’s window less the room kept for its next answer: that room cannot be filled by the conversation. What you are about to send counts too, as you type it and as you add files, before you send it - an estimate, which errs on the high side. A PDF is counted by its pages and its text, read on your device: nothing leaves it before you send.

Past 80 %, a line above the message box says so: This conversation fills 85% of the model’s context. Start a new chat, or switch to a model with a larger window. Once what you are about to send no longer fits, the bar turns red and the line says This conversation no longer fits in the model’s context. Start a new chat, or switch to a model with a larger window.

Stopping an answer

While an answer is being written, the send button becomes Stop. Tap it, or press Escape on a keyboard. What was written so far is kept, marked Stopped.

An answer can also end early without you stopping it. Teloneum then says so under the answer:

The model failed before the end of its answer. - the provider stopped answering midway. What it wrote is kept.
Interrupted before the end of the answer. - the answer was cut on Teloneum’s side. What was written is kept.
The answer reached its maximum length before the end. Ask the model to continue, or to do less at once. - the model wrote, or thought, as much as an answer allows. A model that thinks at length can reach it before it has even made the file you asked for. What was written is kept.

Leaving and coming back

An answer keeps being written even when you are not watching. Reloading the page, closing the tab, locking your phone or losing your connection only stops you from following it: the model carries on, and the answer is saved when it ends.

Open the conversation again and you see the answer where it has got to, and it goes on from there. The address of a conversation names it, so reloading the page brings you back to the same one.

Finding a conversation

The sidebar lists your conversations, most recent first, grouped under Today, Yesterday, Previous 7 days, Previous 30 days and Older. Beside each one, the time it was last written to (today), its day (this week) or its date. A stretch shows its five most recent conversations; Show more under it shows the others. Click the magnifier beside New chat and type in Search chats to narrow the list; click it again to close the search. New chat starts a fresh conversation.

A new conversation shows at the top of the list as soon as you send its first message, as New conversation until it takes the name of your message. From the moment you send a message until its answer is complete, a small pulsing sparkle beside the conversation shows that Teloneum is thinking or writing, even when you have moved to another one.

Answers can be selected and copied.

Deleting a conversation

Open the More menu beside a conversation in the sidebar, choose Delete, then confirm. The conversation and every message in it are gone for good. Deleting a conversation whose answer is still being written stops that answer.

Deleting a conversation does not give back what it counted against your usage limits.

Choosing the language

Teloneum speaks English and French. It opens in the language of the site you came from: click Essayer gratuitement on the French site and it speaks French, Try it free on the English site and it speaks English. Opened directly, it follows your browser’s language.

To change it, click your name at the foot of the sidebar, choose Settings, then pick Français or English under Language. Teloneum remembers the choice on this device. What providers publish about their models - the notes on a certification, the name of a place - is shown as they wrote it.

Choosing light or dark

Teloneum follows your device: light when it is set to light, dark when it is set to dark. To choose for yourself, click the icon beside your name, at the foot of the sidebar, and pick System, Light or Dark - or choose it under Theme in Settings. The icon shows the theme chosen, and Teloneum remembers it on this device.